git.sesse.net Git - ffmpeg/blob - libavcodec/h264.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #define UNCHECKED_BITSTREAM_READER 1
  29
  30 #include "libavutil/imgutils.h"
  31 #include "libavutil/opt.h"
  32 #include "internal.h"
  33 #include "cabac.h"
  34 #include "cabac_functions.h"
  35 #include "dsputil.h"
  36 #include "avcodec.h"
  37 #include "mpegvideo.h"
  38 #include "h264.h"
  39 #include "h264data.h"
  40 #include "h264_mvpred.h"
  41 #include "golomb.h"
  42 #include "mathops.h"
  43 #include "rectangle.h"
  44 #include "thread.h"
  45 #include "vdpau_internal.h"
  46 #include "libavutil/avassert.h"
  47
  48 // #undef NDEBUG
  49 #include <assert.h>
  50
  51 const uint16_t ff_h264_mb_sizes[4] = { 256, 384, 512, 768 };
  52
  53 static const uint8_t rem6[QP_MAX_NUM + 1] = {
  54     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
  55     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
  56     0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2,
  57     3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5,
  58     0, 1, 2, 3,
  59 };
  60
  61 static const uint8_t div6[QP_MAX_NUM + 1] = {
  62     0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3,  3,  3,
  63     3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6,  6,  6,
  64     7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 10, 10, 10,
  65    10,10,10,11,11,11,11,11,11,12,12,12,12,12,12,13,13,13, 13, 13, 13,
  66    14,14,14,14,
  67 };
  68
  69 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  70     PIX_FMT_DXVA2_VLD,
  71     PIX_FMT_VAAPI_VLD,
  72     PIX_FMT_VDA_VLD,
  73     PIX_FMT_YUVJ420P,
  74     PIX_FMT_NONE
  75 };
  76
  77 int avpriv_h264_has_num_reorder_frames(AVCodecContext *avctx)
  78 {
  79     H264Context *h = avctx->priv_data;
  80     return h ? h->sps.num_reorder_frames : 0;
  81 }
  82
  83 /**
  84  * Check if the top & left blocks are available if needed and
  85  * change the dc mode so it only uses the available blocks.
  86  */
  87 int ff_h264_check_intra4x4_pred_mode(H264Context *h)
  88 {
  89     MpegEncContext *const s     = &h->s;
  90     static const int8_t top[12] = {
  91         -1, 0, LEFT_DC_PRED, -1, -1, -1, -1, -1, 0
  92     };
  93     static const int8_t left[12] = {
  94         0, -1, TOP_DC_PRED, 0, -1, -1, -1, 0, -1, DC_128_PRED
  95     };
  96     int i;
  97
  98     if (!(h->top_samples_available & 0x8000)) {
  99         for (i = 0; i < 4; i++) {
 100             int status = top[h->intra4x4_pred_mode_cache[scan8[0] + i]];
 101             if (status < 0) {
 102                 av_log(h->s.avctx, AV_LOG_ERROR,
 103                        "top block unavailable for requested intra4x4 mode %d at %d %d\n",
 104                        status, s->mb_x, s->mb_y);
 105                 return -1;
 106             } else if (status) {
 107                 h->intra4x4_pred_mode_cache[scan8[0] + i] = status;
 108             }
 109         }
 110     }
 111
 112     if ((h->left_samples_available & 0x8888) != 0x8888) {
 113         static const int mask[4] = { 0x8000, 0x2000, 0x80, 0x20 };
 114         for (i = 0; i < 4; i++)
 115             if (!(h->left_samples_available & mask[i])) {
 116                 int status = left[h->intra4x4_pred_mode_cache[scan8[0] + 8 * i]];
 117                 if (status < 0) {
 118                     av_log(h->s.avctx, AV_LOG_ERROR,
 119                            "left block unavailable for requested intra4x4 mode %d at %d %d\n",
 120                            status, s->mb_x, s->mb_y);
 121                     return -1;
 122                 } else if (status) {
 123                     h->intra4x4_pred_mode_cache[scan8[0] + 8 * i] = status;
 124                 }
 125             }
 126     }
 127
 128     return 0;
 129 } // FIXME cleanup like ff_h264_check_intra_pred_mode
 130
 131 /**
 132  * Check if the top & left blocks are available if needed and
 133  * change the dc mode so it only uses the available blocks.
 134  */
 135 int ff_h264_check_intra_pred_mode(H264Context *h, int mode, int is_chroma)
 136 {
 137     MpegEncContext *const s     = &h->s;
 138     static const int8_t top[7]  = { LEFT_DC_PRED8x8, 1, -1, -1 };
 139     static const int8_t left[7] = { TOP_DC_PRED8x8, -1, 2, -1, DC_128_PRED8x8 };
 140
 141     if (mode > 6U) {
 142         av_log(h->s.avctx, AV_LOG_ERROR,
 143                "out of range intra chroma pred mode at %d %d\n",
 144                s->mb_x, s->mb_y);
 145         return -1;
 146     }
 147
 148     if (!(h->top_samples_available & 0x8000)) {
 149         mode = top[mode];
 150         if (mode < 0) {
 151             av_log(h->s.avctx, AV_LOG_ERROR,
 152                    "top block unavailable for requested intra mode at %d %d\n",
 153                    s->mb_x, s->mb_y);
 154             return -1;
 155         }
 156     }
 157
 158     if ((h->left_samples_available & 0x8080) != 0x8080) {
 159         mode = left[mode];
 160         if (is_chroma && (h->left_samples_available & 0x8080)) {
 161             // mad cow disease mode, aka MBAFF + constrained_intra_pred
 162             mode = ALZHEIMER_DC_L0T_PRED8x8 +
 163                    (!(h->left_samples_available & 0x8000)) +
 164                    2 * (mode == DC_128_PRED8x8);
 165         }
 166         if (mode < 0) {
 167             av_log(h->s.avctx, AV_LOG_ERROR,
 168                    "left block unavailable for requested intra mode at %d %d\n",
 169                    s->mb_x, s->mb_y);
 170             return -1;
 171         }
 172     }
 173
 174     return mode;
 175 }
 176
 177 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
 178                                   int *dst_length, int *consumed, int length)
 179 {
 180     int i, si, di;
 181     uint8_t *dst;
 182     int bufidx;
 183
 184     // src[0]&0x80; // forbidden bit
 185     h->nal_ref_idc   = src[0] >> 5;
 186     h->nal_unit_type = src[0] & 0x1F;
 187
 188     src++;
 189     length--;
 190
 191 #define STARTCODE_TEST                                                  \
 192         if (i + 2 < length && src[i + 1] == 0 && src[i + 2] <= 3) {     \
 193             if (src[i + 2] != 3) {                                      \
 194                 /* startcode, so we must be past the end */             \
 195                 length = i;                                             \
 196             }                                                           \
 197             break;                                                      \
 198         }
 199 #if HAVE_FAST_UNALIGNED
 200 #define FIND_FIRST_ZERO                                                 \
 201         if (i > 0 && !src[i])                                           \
 202             i--;                                                        \
 203         while (src[i])                                                  \
 204             i++
 205 #if HAVE_FAST_64BIT
 206     for (i = 0; i + 1 < length; i += 9) {
 207         if (!((~AV_RN64A(src + i) &
 208                (AV_RN64A(src + i) - 0x0100010001000101ULL)) &
 209               0x8000800080008080ULL))
 210             continue;
 211         FIND_FIRST_ZERO;
 212         STARTCODE_TEST;
 213         i -= 7;
 214     }
 215 #else
 216     for (i = 0; i + 1 < length; i += 5) {
 217         if (!((~AV_RN32A(src + i) &
 218                (AV_RN32A(src + i) - 0x01000101U)) &
 219               0x80008080U))
 220             continue;
 221         FIND_FIRST_ZERO;
 222         STARTCODE_TEST;
 223         i -= 3;
 224     }
 225 #endif
 226 #else
 227     for (i = 0; i + 1 < length; i += 2) {
 228         if (src[i])
 229             continue;
 230         if (i > 0 && src[i - 1] == 0)
 231             i--;
 232         STARTCODE_TEST;
 233     }
 234 #endif
 235
 236     // use second escape buffer for inter data
 237     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0;
 238
 239     si = h->rbsp_buffer_size[bufidx];
 240     av_fast_padded_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+MAX_MBPAIR_SIZE);
 241     dst = h->rbsp_buffer[bufidx];
 242
 243     if (dst == NULL)
 244         return NULL;
 245
 246     if(i>=length-1){ //no escaped 0
 247         *dst_length= length;
 248         *consumed= length+1; //+1 for the header
 249         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
 250             return src;
 251         }else{
 252             memcpy(dst, src, length);
 253             return dst;
 254         }
 255     }
 256
 257     // printf("decoding esc\n");
 258     memcpy(dst, src, i);
 259     si = di = i;
 260     while (si + 2 < length) {
 261         // remove escapes (very rare 1:2^22)
 262         if (src[si + 2] > 3) {
 263             dst[di++] = src[si++];
 264             dst[di++] = src[si++];
 265         } else if (src[si] == 0 && src[si + 1] == 0) {
 266             if (src[si + 2] == 3) { // escape
 267                 dst[di++]  = 0;
 268                 dst[di++]  = 0;
 269                 si        += 3;
 270                 continue;
 271             } else // next start code
 272                 goto nsc;
 273         }
 274
 275         dst[di++] = src[si++];
 276     }
 277     while (si < length)
 278         dst[di++] = src[si++];
 279 nsc:
 280
 281     memset(dst + di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 282
 283     *dst_length = di;
 284     *consumed   = si + 1; // +1 for the header
 285     /* FIXME store exact number of bits in the getbitcontext
 286      * (it is needed for decoding) */
 287     return dst;
 288 }
 289
 290 /**
 291  * Identify the exact end of the bitstream
 292  * @return the length of the trailing, or 0 if damaged
 293  */
 294 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src)
 295 {
 296     int v = *src;
 297     int r;
 298
 299     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 300
 301     for (r = 1; r < 9; r++) {
 302         if (v & 1)
 303             return r;
 304         v >>= 1;
 305     }
 306     return 0;
 307 }
 308
 309 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n,
 310                                          int height, int y_offset, int list)
 311 {
 312     int raw_my        = h->mv_cache[list][scan8[n]][1];
 313     int filter_height = (raw_my & 3) ? 2 : 0;
 314     int full_my       = (raw_my >> 2) + y_offset;
 315     int top           = full_my - filter_height;
 316     int bottom        = full_my + filter_height + height;
 317
 318     return FFMAX(abs(top), bottom);
 319 }
 320
 321 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
 322                                      int height, int y_offset, int list0,
 323                                      int list1, int *nrefs)
 324 {
 325     MpegEncContext *const s = &h->s;
 326     int my;
 327
 328     y_offset += 16 * (s->mb_y >> MB_FIELD);
 329
 330     if (list0) {
 331         int ref_n    = h->ref_cache[0][scan8[n]];
 332         Picture *ref = &h->ref_list[0][ref_n];
 333
 334         // Error resilience puts the current picture in the ref list.
 335         // Don't try to wait on these as it will cause a deadlock.
 336         // Fields can wait on each other, though.
 337         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 338             (ref->f.reference & 3) != s->picture_structure) {
 339             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 340             if (refs[0][ref_n] < 0)
 341                 nrefs[0] += 1;
 342             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 343         }
 344     }
 345
 346     if (list1) {
 347         int ref_n    = h->ref_cache[1][scan8[n]];
 348         Picture *ref = &h->ref_list[1][ref_n];
 349
 350         if (ref->f.thread_opaque   != s->current_picture.f.thread_opaque ||
 351             (ref->f.reference & 3) != s->picture_structure) {
 352             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 353             if (refs[1][ref_n] < 0)
 354                 nrefs[1] += 1;
 355             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 356         }
 357     }
 358 }
 359
 360 /**
 361  * Wait until all reference frames are available for MC operations.
 362  *
 363  * @param h the H264 context
 364  */
 365 static void await_references(H264Context *h)
 366 {
 367     MpegEncContext *const s = &h->s;
 368     const int mb_xy   = h->mb_xy;
 369     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 370     int refs[2][48];
 371     int nrefs[2] = { 0 };
 372     int ref, list;
 373
 374     memset(refs, -1, sizeof(refs));
 375
 376     if (IS_16X16(mb_type)) {
 377         get_lowest_part_y(h, refs, 0, 16, 0,
 378                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 379     } else if (IS_16X8(mb_type)) {
 380         get_lowest_part_y(h, refs, 0, 8, 0,
 381                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 382         get_lowest_part_y(h, refs, 8, 8, 8,
 383                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 384     } else if (IS_8X16(mb_type)) {
 385         get_lowest_part_y(h, refs, 0, 16, 0,
 386                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 387         get_lowest_part_y(h, refs, 4, 16, 0,
 388                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 389     } else {
 390         int i;
 391
 392         assert(IS_8X8(mb_type));
 393
 394         for (i = 0; i < 4; i++) {
 395             const int sub_mb_type = h->sub_mb_type[i];
 396             const int n           = 4 * i;
 397             int y_offset          = (i & 2) << 2;
 398
 399             if (IS_SUB_8X8(sub_mb_type)) {
 400                 get_lowest_part_y(h, refs, n, 8, y_offset,
 401                                   IS_DIR(sub_mb_type, 0, 0),
 402                                   IS_DIR(sub_mb_type, 0, 1),
 403                                   nrefs);
 404             } else if (IS_SUB_8X4(sub_mb_type)) {
 405                 get_lowest_part_y(h, refs, n, 4, y_offset,
 406                                   IS_DIR(sub_mb_type, 0, 0),
 407                                   IS_DIR(sub_mb_type, 0, 1),
 408                                   nrefs);
 409                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
 410                                   IS_DIR(sub_mb_type, 0, 0),
 411                                   IS_DIR(sub_mb_type, 0, 1),
 412                                   nrefs);
 413             } else if (IS_SUB_4X8(sub_mb_type)) {
 414                 get_lowest_part_y(h, refs, n, 8, y_offset,
 415                                   IS_DIR(sub_mb_type, 0, 0),
 416                                   IS_DIR(sub_mb_type, 0, 1),
 417                                   nrefs);
 418                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
 419                                   IS_DIR(sub_mb_type, 0, 0),
 420                                   IS_DIR(sub_mb_type, 0, 1),
 421                                   nrefs);
 422             } else {
 423                 int j;
 424                 assert(IS_SUB_4X4(sub_mb_type));
 425                 for (j = 0; j < 4; j++) {
 426                     int sub_y_offset = y_offset + 2 * (j & 2);
 427                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
 428                                       IS_DIR(sub_mb_type, 0, 0),
 429                                       IS_DIR(sub_mb_type, 0, 1),
 430                                       nrefs);
 431                 }
 432             }
 433         }
 434     }
 435
 436     for (list = h->list_count - 1; list >= 0; list--)
 437         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
 438             int row = refs[list][ref];
 439             if (row >= 0) {
 440                 Picture *ref_pic      = &h->ref_list[list][ref];
 441                 int ref_field         = ref_pic->f.reference - 1;
 442                 int ref_field_picture = ref_pic->field_picture;
 443                 int pic_height        = 16 * s->mb_height >> ref_field_picture;
 444
 445                 row <<= MB_MBAFF;
 446                 nrefs[list]--;
 447
 448                 if (!FIELD_PICTURE && ref_field_picture) { // frame referencing two fields
 449                     ff_thread_await_progress(&ref_pic->f,
 450                                              FFMIN((row >> 1) - !(row & 1),
 451                                                    pic_height - 1),
 452                                              1);
 453                     ff_thread_await_progress(&ref_pic->f,
 454                                              FFMIN((row >> 1), pic_height - 1),
 455                                              0);
 456                 } else if (FIELD_PICTURE && !ref_field_picture) { // field referencing one field of a frame
 457                     ff_thread_await_progress(&ref_pic->f,
 458                                              FFMIN(row * 2 + ref_field,
 459                                                    pic_height - 1),
 460                                              0);
 461                 } else if (FIELD_PICTURE) {
 462                     ff_thread_await_progress(&ref_pic->f,
 463                                              FFMIN(row, pic_height - 1),
 464                                              ref_field);
 465                 } else {
 466                     ff_thread_await_progress(&ref_pic->f,
 467                                              FFMIN(row, pic_height - 1),
 468                                              0);
 469                 }
 470             }
 471         }
 472 }
 473
 474 static av_always_inline void mc_dir_part(H264Context *h, Picture *pic,
 475                                          int n, int square, int height,
 476                                          int delta, int list,
 477                                          uint8_t *dest_y, uint8_t *dest_cb,
 478                                          uint8_t *dest_cr,
 479                                          int src_x_offset, int src_y_offset,
 480                                          qpel_mc_func *qpix_op,
 481                                          h264_chroma_mc_func chroma_op,
 482                                          int pixel_shift, int chroma_idc)
 483 {
 484     MpegEncContext *const s = &h->s;
 485     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
 486     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
 487     const int luma_xy = (mx & 3) + ((my & 3) << 2);
 488     int offset        = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
 489     uint8_t *src_y    = pic->f.data[0] + offset;
 490     uint8_t *src_cb, *src_cr;
 491     int extra_width  = h->emu_edge_width;
 492     int extra_height = h->emu_edge_height;
 493     int emu = 0;
 494     const int full_mx    = mx >> 2;
 495     const int full_my    = my >> 2;
 496     const int pic_width  = 16 * s->mb_width;
 497     const int pic_height = 16 * s->mb_height >> MB_FIELD;
 498     int ysh;
 499
 500     if (mx & 7)
 501         extra_width -= 3;
 502     if (my & 7)
 503         extra_height -= 3;
 504
 505     if (full_mx                <          0 - extra_width  ||
 506         full_my                <          0 - extra_height ||
 507         full_mx + 16 /*FIXME*/ > pic_width  + extra_width  ||
 508         full_my + 16 /*FIXME*/ > pic_height + extra_height) {
 509         s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 510                                 src_y - (2 << pixel_shift) - 2 * h->mb_linesize,
 511                                 h->mb_linesize,
 512                                 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
 513                                 full_my - 2, pic_width, pic_height);
 514         src_y = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 515         emu   = 1;
 516     }
 517
 518     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
 519     if (!square)
 520         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 521
 522     if (CONFIG_GRAY && s->flags & CODEC_FLAG_GRAY)
 523         return;
 524
 525     if (chroma_idc == 3 /* yuv444 */) {
 526         src_cb = pic->f.data[1] + offset;
 527         if (emu) {
 528             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 529                                     src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
 530                                     h->mb_linesize,
 531                                     16 + 5, 16 + 5 /*FIXME*/,
 532                                     full_mx - 2, full_my - 2,
 533                                     pic_width, pic_height);
 534             src_cb = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 535         }
 536         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
 537         if (!square)
 538             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 539
 540         src_cr = pic->f.data[2] + offset;
 541         if (emu) {
 542             s->dsp.emulated_edge_mc(s->edge_emu_buffer,
 543                                     src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
 544                                     h->mb_linesize,
 545                                     16 + 5, 16 + 5 /*FIXME*/,
 546                                     full_mx - 2, full_my - 2,
 547                                     pic_width, pic_height);
 548             src_cr = s->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
 549         }
 550         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
 551         if (!square)
 552             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 553         return;
 554     }
 555
 556     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 557     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD) {
 558         // chroma offset when predicting from a field of opposite parity
 559         my  += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 560         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
 561     }
 562
 563     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
 564              (my >> ysh) * h->mb_uvlinesize;
 565     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
 566              (my >> ysh) * h->mb_uvlinesize;
 567
 568     if (emu) {
 569         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 570                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 571                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 572         src_cb = s->edge_emu_buffer;
 573     }
 574     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
 575               height >> (chroma_idc == 1 /* yuv420 */),
 576               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 577
 578     if (emu) {
 579         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 580                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 581                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 582         src_cr = s->edge_emu_buffer;
 583     }
 584     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 585               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
 586 }
 587
 588 static av_always_inline void mc_part_std(H264Context *h, int n, int square,
 589                                          int height, int delta,
 590                                          uint8_t *dest_y, uint8_t *dest_cb,
 591                                          uint8_t *dest_cr,
 592                                          int x_offset, int y_offset,
 593                                          qpel_mc_func *qpix_put,
 594                                          h264_chroma_mc_func chroma_put,
 595                                          qpel_mc_func *qpix_avg,
 596                                          h264_chroma_mc_func chroma_avg,
 597                                          int list0, int list1,
 598                                          int pixel_shift, int chroma_idc)
 599 {
 600     MpegEncContext *const s       = &h->s;
 601     qpel_mc_func *qpix_op         = qpix_put;
 602     h264_chroma_mc_func chroma_op = chroma_put;
 603
 604     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 605     if (chroma_idc == 3 /* yuv444 */) {
 606         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 607         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 608     } else if (chroma_idc == 2 /* yuv422 */) {
 609         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 610         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 611     } else { /* yuv420 */
 612         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 613         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 614     }
 615     x_offset += 8 * s->mb_x;
 616     y_offset += 8 * (s->mb_y >> MB_FIELD);
 617
 618     if (list0) {
 619         Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
 620         mc_dir_part(h, ref, n, square, height, delta, 0,
 621                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 622                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 623
 624         qpix_op   = qpix_avg;
 625         chroma_op = chroma_avg;
 626     }
 627
 628     if (list1) {
 629         Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
 630         mc_dir_part(h, ref, n, square, height, delta, 1,
 631                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 632                     qpix_op, chroma_op, pixel_shift, chroma_idc);
 633     }
 634 }
 635
 636 static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
 637                                               int height, int delta,
 638                                               uint8_t *dest_y, uint8_t *dest_cb,
 639                                               uint8_t *dest_cr,
 640                                               int x_offset, int y_offset,
 641                                               qpel_mc_func *qpix_put,
 642                                               h264_chroma_mc_func chroma_put,
 643                                               h264_weight_func luma_weight_op,
 644                                               h264_weight_func chroma_weight_op,
 645                                               h264_biweight_func luma_weight_avg,
 646                                               h264_biweight_func chroma_weight_avg,
 647                                               int list0, int list1,
 648                                               int pixel_shift, int chroma_idc)
 649 {
 650     MpegEncContext *const s = &h->s;
 651     int chroma_height;
 652
 653     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 654     if (chroma_idc == 3 /* yuv444 */) {
 655         chroma_height     = height;
 656         chroma_weight_avg = luma_weight_avg;
 657         chroma_weight_op  = luma_weight_op;
 658         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 659         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
 660     } else if (chroma_idc == 2 /* yuv422 */) {
 661         chroma_height = height;
 662         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 663         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
 664     } else { /* yuv420 */
 665         chroma_height = height >> 1;
 666         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 667         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
 668     }
 669     x_offset += 8 * s->mb_x;
 670     y_offset += 8 * (s->mb_y >> MB_FIELD);
 671
 672     if (list0 && list1) {
 673         /* don't optimize for luma-only case, since B-frames usually
 674          * use implicit weights => chroma too. */
 675         uint8_t *tmp_cb = s->obmc_scratchpad;
 676         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 677         uint8_t *tmp_y  = s->obmc_scratchpad + 16 * h->mb_uvlinesize;
 678         int refn0       = h->ref_cache[0][scan8[n]];
 679         int refn1       = h->ref_cache[1][scan8[n]];
 680
 681         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 682                     dest_y, dest_cb, dest_cr,
 683                     x_offset, y_offset, qpix_put, chroma_put,
 684                     pixel_shift, chroma_idc);
 685         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 686                     tmp_y, tmp_cb, tmp_cr,
 687                     x_offset, y_offset, qpix_put, chroma_put,
 688                     pixel_shift, chroma_idc);
 689
 690         if (h->use_weight == 2) {
 691             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y & 1];
 692             int weight1 = 64 - weight0;
 693             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
 694                             height, 5, weight0, weight1, 0);
 695             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 696                               chroma_height, 5, weight0, weight1, 0);
 697             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 698                               chroma_height, 5, weight0, weight1, 0);
 699         } else {
 700             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
 701                             h->luma_log2_weight_denom,
 702                             h->luma_weight[refn0][0][0],
 703                             h->luma_weight[refn1][1][0],
 704                             h->luma_weight[refn0][0][1] +
 705                             h->luma_weight[refn1][1][1]);
 706             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
 707                               h->chroma_log2_weight_denom,
 708                               h->chroma_weight[refn0][0][0][0],
 709                               h->chroma_weight[refn1][1][0][0],
 710                               h->chroma_weight[refn0][0][0][1] +
 711                               h->chroma_weight[refn1][1][0][1]);
 712             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
 713                               h->chroma_log2_weight_denom,
 714                               h->chroma_weight[refn0][0][1][0],
 715                               h->chroma_weight[refn1][1][1][0],
 716                               h->chroma_weight[refn0][0][1][1] +
 717                               h->chroma_weight[refn1][1][1][1]);
 718         }
 719     } else {
 720         int list     = list1 ? 1 : 0;
 721         int refn     = h->ref_cache[list][scan8[n]];
 722         Picture *ref = &h->ref_list[list][refn];
 723         mc_dir_part(h, ref, n, square, height, delta, list,
 724                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 725                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 726
 727         luma_weight_op(dest_y, h->mb_linesize, height,
 728                        h->luma_log2_weight_denom,
 729                        h->luma_weight[refn][list][0],
 730                        h->luma_weight[refn][list][1]);
 731         if (h->use_weight_chroma) {
 732             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
 733                              h->chroma_log2_weight_denom,
 734                              h->chroma_weight[refn][list][0][0],
 735                              h->chroma_weight[refn][list][0][1]);
 736             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
 737                              h->chroma_log2_weight_denom,
 738                              h->chroma_weight[refn][list][1][0],
 739                              h->chroma_weight[refn][list][1][1]);
 740         }
 741     }
 742 }
 743
 744 static av_always_inline void prefetch_motion(H264Context *h, int list,
 745                                              int pixel_shift, int chroma_idc)
 746 {
 747     /* fetch pixels for estimated mv 4 macroblocks ahead
 748      * optimized for 64byte cache lines */
 749     MpegEncContext *const s = &h->s;
 750     const int refn = h->ref_cache[list][scan8[0]];
 751     if (refn >= 0) {
 752         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * s->mb_x + 8;
 753         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * s->mb_y;
 754         uint8_t **src = h->ref_list[list][refn].f.data;
 755         int off       = (mx << pixel_shift) +
 756                         (my + (s->mb_x & 3) * 4) * h->mb_linesize +
 757                         (64 << pixel_shift);
 758         s->dsp.prefetch(src[0] + off, s->linesize, 4);
 759         if (chroma_idc == 3 /* yuv444 */) {
 760             s->dsp.prefetch(src[1] + off, s->linesize, 4);
 761             s->dsp.prefetch(src[2] + off, s->linesize, 4);
 762         } else {
 763             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 764             s->dsp.prefetch(src[1] + off, src[2] - src[1], 2);
 765         }
 766     }
 767 }
 768
 769 static void free_tables(H264Context *h, int free_rbsp)
 770 {
 771     int i;
 772     H264Context *hx;
 773
 774     av_freep(&h->intra4x4_pred_mode);
 775     av_freep(&h->chroma_pred_mode_table);
 776     av_freep(&h->cbp_table);
 777     av_freep(&h->mvd_table[0]);
 778     av_freep(&h->mvd_table[1]);
 779     av_freep(&h->direct_table);
 780     av_freep(&h->non_zero_count);
 781     av_freep(&h->slice_table_base);
 782     h->slice_table = NULL;
 783     av_freep(&h->list_counts);
 784
 785     av_freep(&h->mb2b_xy);
 786     av_freep(&h->mb2br_xy);
 787
 788     for (i = 0; i < MAX_THREADS; i++) {
 789         hx = h->thread_context[i];
 790         if (!hx)
 791             continue;
 792         av_freep(&hx->top_borders[1]);
 793         av_freep(&hx->top_borders[0]);
 794         av_freep(&hx->s.obmc_scratchpad);
 795         if (free_rbsp) {
 796             av_freep(&hx->rbsp_buffer[1]);
 797             av_freep(&hx->rbsp_buffer[0]);
 798             hx->rbsp_buffer_size[0] = 0;
 799             hx->rbsp_buffer_size[1] = 0;
 800         }
 801         if (i)
 802             av_freep(&h->thread_context[i]);
 803     }
 804 }
 805
 806 static void init_dequant8_coeff_table(H264Context *h)
 807 {
 808     int i, j, q, x;
 809     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 810
 811     for (i = 0; i < 6; i++) {
 812         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 813         for (j = 0; j < i; j++)
 814             if (!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i],
 815                         64 * sizeof(uint8_t))) {
 816                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 817                 break;
 818             }
 819         if (j < i)
 820             continue;
 821
 822         for (q = 0; q < max_qp + 1; q++) {
 823             int shift = div6[q];
 824             int idx   = rem6[q];
 825             for (x = 0; x < 64; x++)
 826                 h->dequant8_coeff[i][q][(x >> 3) | ((x & 7) << 3)] =
 827                     ((uint32_t)dequant8_coeff_init[idx][dequant8_coeff_init_scan[((x >> 1) & 12) | (x & 3)]] *
 828                      h->pps.scaling_matrix8[i][x]) << shift;
 829         }
 830     }
 831 }
 832
 833 static void init_dequant4_coeff_table(H264Context *h)
 834 {
 835     int i, j, q, x;
 836     const int max_qp = 51 + 6 * (h->sps.bit_depth_luma - 8);
 837     for (i = 0; i < 6; i++) {
 838         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 839         for (j = 0; j < i; j++)
 840             if (!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i],
 841                         16 * sizeof(uint8_t))) {
 842                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 843                 break;
 844             }
 845         if (j < i)
 846             continue;
 847
 848         for (q = 0; q < max_qp + 1; q++) {
 849             int shift = div6[q] + 2;
 850             int idx   = rem6[q];
 851             for (x = 0; x < 16; x++)
 852                 h->dequant4_coeff[i][q][(x >> 2) | ((x << 2) & 0xF)] =
 853                     ((uint32_t)dequant4_coeff_init[idx][(x & 1) + ((x >> 2) & 1)] *
 854                      h->pps.scaling_matrix4[i][x]) << shift;
 855         }
 856     }
 857 }
 858
 859 static void init_dequant_tables(H264Context *h)
 860 {
 861     int i, x;
 862     init_dequant4_coeff_table(h);
 863     if (h->pps.transform_8x8_mode)
 864         init_dequant8_coeff_table(h);
 865     if (h->sps.transform_bypass) {
 866         for (i = 0; i < 6; i++)
 867             for (x = 0; x < 16; x++)
 868                 h->dequant4_coeff[i][0][x] = 1 << 6;
 869         if (h->pps.transform_8x8_mode)
 870             for (i = 0; i < 6; i++)
 871                 for (x = 0; x < 64; x++)
 872                     h->dequant8_coeff[i][0][x] = 1 << 6;
 873     }
 874 }
 875
 876 int ff_h264_alloc_tables(H264Context *h)
 877 {
 878     MpegEncContext *const s = &h->s;
 879     const int big_mb_num    = s->mb_stride * (s->mb_height + 1);
 880     const int row_mb_num    = 2*s->mb_stride*FFMAX(s->avctx->thread_count, 1);
 881     int x, y;
 882
 883     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode,
 884                       row_mb_num * 8 * sizeof(uint8_t), fail)
 885     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count,
 886                       big_mb_num * 48 * sizeof(uint8_t), fail)
 887     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base,
 888                       (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base), fail)
 889     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table,
 890                       big_mb_num * sizeof(uint16_t), fail)
 891     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table,
 892                       big_mb_num * sizeof(uint8_t), fail)
 893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0],
 894                       16 * row_mb_num * sizeof(uint8_t), fail);
 895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1],
 896                       16 * row_mb_num * sizeof(uint8_t), fail);
 897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table,
 898                       4 * big_mb_num * sizeof(uint8_t), fail);
 899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts,
 900                       big_mb_num * sizeof(uint8_t), fail)
 901
 902     memset(h->slice_table_base, -1,
 903            (big_mb_num + s->mb_stride) * sizeof(*h->slice_table_base));
 904     h->slice_table = h->slice_table_base + s->mb_stride * 2 + 1;
 905
 906     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy,
 907                       big_mb_num * sizeof(uint32_t), fail);
 908     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy,
 909                       big_mb_num * sizeof(uint32_t), fail);
 910     for (y = 0; y < s->mb_height; y++)
 911         for (x = 0; x < s->mb_width; x++) {
 912             const int mb_xy = x + y * s->mb_stride;
 913             const int b_xy  = 4 * x + 4 * y * h->b_stride;
 914
 915             h->mb2b_xy[mb_xy]  = b_xy;
 916             h->mb2br_xy[mb_xy] = 8 * (FMO ? mb_xy : (mb_xy % (2 * s->mb_stride)));
 917         }
 918
 919     s->obmc_scratchpad = NULL;
 920
 921     if (!h->dequant4_coeff[0])
 922         init_dequant_tables(h);
 923
 924     return 0;
 925
 926 fail:
 927     free_tables(h, 1);
 928     return -1;
 929 }
 930
 931 /**
 932  * Mimic alloc_tables(), but for every context thread.
 933  */
 934 static void clone_tables(H264Context *dst, H264Context *src, int i)
 935 {
 936     MpegEncContext *const s     = &src->s;
 937     dst->intra4x4_pred_mode     = src->intra4x4_pred_mode + i * 8 * 2 * s->mb_stride;
 938     dst->non_zero_count         = src->non_zero_count;
 939     dst->slice_table            = src->slice_table;
 940     dst->cbp_table              = src->cbp_table;
 941     dst->mb2b_xy                = src->mb2b_xy;
 942     dst->mb2br_xy               = src->mb2br_xy;
 943     dst->chroma_pred_mode_table = src->chroma_pred_mode_table;
 944     dst->mvd_table[0]           = src->mvd_table[0] + i * 8 * 2 * s->mb_stride;
 945     dst->mvd_table[1]           = src->mvd_table[1] + i * 8 * 2 * s->mb_stride;
 946     dst->direct_table           = src->direct_table;
 947     dst->list_counts            = src->list_counts;
 948     dst->s.obmc_scratchpad      = NULL;
 949     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma,
 950                       src->sps.chroma_format_idc);
 951 }
 952
 953 /**
 954  * Init context
 955  * Allocate buffers which are not shared amongst multiple threads.
 956  */
 957 static int context_init(H264Context *h)
 958 {
 959     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0],
 960                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
 961     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1],
 962                       h->s.mb_width * 16 * 3 * sizeof(uint8_t) * 2, fail)
 963
 964     h->ref_cache[0][scan8[5]  + 1] =
 965     h->ref_cache[0][scan8[7]  + 1] =
 966     h->ref_cache[0][scan8[13] + 1] =
 967     h->ref_cache[1][scan8[5]  + 1] =
 968     h->ref_cache[1][scan8[7]  + 1] =
 969     h->ref_cache[1][scan8[13] + 1] = PART_NOT_AVAILABLE;
 970
 971     return 0;
 972
 973 fail:
 974     return -1; // free_tables will clean up for us
 975 }
 976
 977 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 978
 979 static av_cold void common_init(H264Context *h)
 980 {
 981     MpegEncContext *const s = &h->s;
 982
 983     s->width    = s->avctx->width;
 984     s->height   = s->avctx->height;
 985     s->codec_id = s->avctx->codec->id;
 986
 987     s->avctx->bits_per_raw_sample = 8;
 988     h->cur_chroma_format_idc = 1;
 989
 990     ff_h264dsp_init(&h->h264dsp,
 991                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
 992     ff_h264_pred_init(&h->hpc, s->codec_id,
 993                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
 994
 995     h->dequant_coeff_pps = -1;
 996     s->unrestricted_mv   = 1;
 997
 998     s->dsp.dct_bits = 16;
 999     /* needed so that IDCT permutation is known early */
1000     ff_dsputil_init(&s->dsp, s->avctx);
1001
1002     memset(h->pps.scaling_matrix4, 16, 6 * 16 * sizeof(uint8_t));
1003     memset(h->pps.scaling_matrix8, 16, 2 * 64 * sizeof(uint8_t));
1004 }
1005
1006 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1007 {
1008     AVCodecContext *avctx = h->s.avctx;
1009
1010     if (!buf || size <= 0)
1011         return -1;
1012
1013     if (buf[0] == 1) {
1014         int i, cnt, nalsize;
1015         const unsigned char *p = buf;
1016
1017         h->is_avc = 1;
1018
1019         if (size < 7) {
1020             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1021             return -1;
1022         }
1023         /* sps and pps in the avcC always have length coded with 2 bytes,
1024          * so put a fake nal_length_size = 2 while parsing them */
1025         h->nal_length_size = 2;
1026         // Decode sps from avcC
1027         cnt = *(p + 5) & 0x1f; // Number of sps
1028         p  += 6;
1029         for (i = 0; i < cnt; i++) {
1030             nalsize = AV_RB16(p) + 2;
1031             if(nalsize > size - (p-buf))
1032                 return -1;
1033             if (decode_nal_units(h, p, nalsize) < 0) {
1034                 av_log(avctx, AV_LOG_ERROR,
1035                        "Decoding sps %d from avcC failed\n", i);
1036                 return -1;
1037             }
1038             p += nalsize;
1039         }
1040         // Decode pps from avcC
1041         cnt = *(p++); // Number of pps
1042         for (i = 0; i < cnt; i++) {
1043             nalsize = AV_RB16(p) + 2;
1044             if(nalsize > size - (p-buf))
1045                 return -1;
1046             if (decode_nal_units(h, p, nalsize) < 0) {
1047                 av_log(avctx, AV_LOG_ERROR,
1048                        "Decoding pps %d from avcC failed\n", i);
1049                 return -1;
1050             }
1051             p += nalsize;
1052         }
1053         // Now store right nal length size, that will be used to parse all other nals
1054         h->nal_length_size = (buf[4] & 0x03) + 1;
1055     } else {
1056         h->is_avc = 0;
1057         if (decode_nal_units(h, buf, size) < 0)
1058             return -1;
1059     }
1060     return size;
1061 }
1062
1063 av_cold int ff_h264_decode_init(AVCodecContext *avctx)
1064 {
1065     H264Context *h = avctx->priv_data;
1066     MpegEncContext *const s = &h->s;
1067     int i;
1068
1069     ff_MPV_decode_defaults(s);
1070
1071     s->avctx = avctx;
1072     common_init(h);
1073
1074     s->out_format      = FMT_H264;
1075     s->workaround_bugs = avctx->workaround_bugs;
1076
1077     /* set defaults */
1078     // s->decode_mb = ff_h263_decode_mb;
1079     s->quarter_sample = 1;
1080     if (!avctx->has_b_frames)
1081         s->low_delay = 1;
1082
1083     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1084
1085     ff_h264_decode_init_vlc();
1086
1087     h->pixel_shift = 0;
1088     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1089
1090     h->thread_context[0] = h;
1091     h->outputed_poc      = h->next_outputed_poc = INT_MIN;
1092     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1093         h->last_pocs[i] = INT_MIN;
1094     h->prev_poc_msb = 1 << 16;
1095     h->prev_frame_num = -1;
1096     h->x264_build   = -1;
1097     ff_h264_reset_sei(h);
1098     if (avctx->codec_id == CODEC_ID_H264) {
1099         if (avctx->ticks_per_frame == 1)
1100             s->avctx->time_base.den *= 2;
1101         avctx->ticks_per_frame = 2;
1102     }
1103
1104     if (avctx->extradata_size > 0 && avctx->extradata &&
1105         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size) < 0) {
1106         ff_h264_free_context(h);
1107         return -1;
1108     }
1109
1110     if (h->sps.bitstream_restriction_flag &&
1111         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1112         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1113         s->low_delay           = 0;
1114     }
1115
1116     return 0;
1117 }
1118
1119 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b) + (size))))
1120
1121 static void copy_picture_range(Picture **to, Picture **from, int count,
1122                                MpegEncContext *new_base,
1123                                MpegEncContext *old_base)
1124 {
1125     int i;
1126
1127     for (i = 0; i < count; i++) {
1128         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1129                 IN_RANGE(from[i], old_base->picture,
1130                          sizeof(Picture) * old_base->picture_count) ||
1131                 !from[i]));
1132         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1133     }
1134 }
1135
1136 static void copy_parameter_set(void **to, void **from, int count, int size)
1137 {
1138     int i;
1139
1140     for (i = 0; i < count; i++) {
1141         if (to[i] && !from[i])
1142             av_freep(&to[i]);
1143         else if (from[i] && !to[i])
1144             to[i] = av_malloc(size);
1145
1146         if (from[i])
1147             memcpy(to[i], from[i], size);
1148     }
1149 }
1150
1151 static int decode_init_thread_copy(AVCodecContext *avctx)
1152 {
1153     H264Context *h = avctx->priv_data;
1154
1155     if (!avctx->internal->is_copy)
1156         return 0;
1157     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1158     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1159
1160     return 0;
1161 }
1162
1163 #define copy_fields(to, from, start_field, end_field)                   \
1164     memcpy(&to->start_field, &from->start_field,                        \
1165            (char *)&to->end_field - (char *)&to->start_field)
1166
1167 static int decode_update_thread_context(AVCodecContext *dst,
1168                                         const AVCodecContext *src)
1169 {
1170     H264Context *h = dst->priv_data, *h1 = src->priv_data;
1171     MpegEncContext *const s = &h->s, *const s1 = &h1->s;
1172     int inited = s->context_initialized, err;
1173     int i;
1174
1175     if (dst == src)
1176         return 0;
1177
1178     err = ff_mpeg_update_thread_context(dst, src);
1179     if (err)
1180         return err;
1181
1182     // FIXME handle width/height changing
1183     if (!inited) {
1184         for (i = 0; i < MAX_SPS_COUNT; i++)
1185             av_freep(h->sps_buffers + i);
1186
1187         for (i = 0; i < MAX_PPS_COUNT; i++)
1188             av_freep(h->pps_buffers + i);
1189
1190         // copy all fields after MpegEnc
1191         memcpy(&h->s + 1, &h1->s + 1,
1192                sizeof(H264Context) - sizeof(MpegEncContext));
1193         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1194         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1195
1196         if (s1->context_initialized) {
1197         if (ff_h264_alloc_tables(h) < 0) {
1198             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1199             return AVERROR(ENOMEM);
1200         }
1201         context_init(h);
1202
1203         /* frame_start may not be called for the next thread (if it's decoding
1204          * a bottom field) so this has to be allocated here */
1205         h->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1206         }
1207
1208         for (i = 0; i < 2; i++) {
1209             h->rbsp_buffer[i]      = NULL;
1210             h->rbsp_buffer_size[i] = 0;
1211         }
1212
1213         h->thread_context[0] = h;
1214
1215         s->dsp.clear_blocks(h->mb);
1216         s->dsp.clear_blocks(h->mb + (24 * 16 << h->pixel_shift));
1217     }
1218
1219     // extradata/NAL handling
1220     h->is_avc = h1->is_avc;
1221
1222     // SPS/PPS
1223     copy_parameter_set((void **)h->sps_buffers, (void **)h1->sps_buffers,
1224                        MAX_SPS_COUNT, sizeof(SPS));
1225     h->sps = h1->sps;
1226     copy_parameter_set((void **)h->pps_buffers, (void **)h1->pps_buffers,
1227                        MAX_PPS_COUNT, sizeof(PPS));
1228     h->pps = h1->pps;
1229
1230     // Dequantization matrices
1231     // FIXME these are big - can they be only copied when PPS changes?
1232     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1233
1234     for (i = 0; i < 6; i++)
1235         h->dequant4_coeff[i] = h->dequant4_buffer[0] +
1236                                (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1237
1238     for (i = 0; i < 6; i++)
1239         h->dequant8_coeff[i] = h->dequant8_buffer[0] +
1240                                (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1241
1242     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1243
1244     // POC timing
1245     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1246
1247     // reference lists
1248     copy_fields(h, h1, ref_count, list_count);
1249     copy_fields(h, h1, ref_list, intra_gb);
1250     copy_fields(h, h1, short_ref, cabac_init_idc);
1251
1252     copy_picture_range(h->short_ref, h1->short_ref, 32, s, s1);
1253     copy_picture_range(h->long_ref, h1->long_ref, 32, s, s1);
1254     copy_picture_range(h->delayed_pic, h1->delayed_pic,
1255                        MAX_DELAYED_PIC_COUNT + 2, s, s1);
1256
1257     h->last_slice_type = h1->last_slice_type;
1258     h->sync            = h1->sync;
1259
1260     if (!s->current_picture_ptr)
1261         return 0;
1262
1263     if (!s->dropable) {
1264         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1265         h->prev_poc_msb = h->poc_msb;
1266         h->prev_poc_lsb = h->poc_lsb;
1267     }
1268     h->prev_frame_num_offset = h->frame_num_offset;
1269     h->prev_frame_num        = h->frame_num;
1270     h->outputed_poc          = h->next_outputed_poc;
1271
1272     return err;
1273 }
1274
1275 int ff_h264_frame_start(H264Context *h)
1276 {
1277     MpegEncContext *const s = &h->s;
1278     int i;
1279     const int pixel_shift = h->pixel_shift;
1280
1281     if (ff_MPV_frame_start(s, s->avctx) < 0)
1282         return -1;
1283     ff_er_frame_start(s);
1284     /*
1285      * ff_MPV_frame_start uses pict_type to derive key_frame.
1286      * This is incorrect for H.264; IDR markings must be used.
1287      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1288      * See decode_nal_units().
1289      */
1290     s->current_picture_ptr->f.key_frame = 0;
1291     s->current_picture_ptr->sync        = 0;
1292     s->current_picture_ptr->mmco_reset  = 0;
1293
1294     assert(s->linesize && s->uvlinesize);
1295
1296     for (i = 0; i < 16; i++) {
1297         h->block_offset[i]           = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1298         h->block_offset[48 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->linesize * ((scan8[i] - scan8[0]) >> 3);
1299     }
1300     for (i = 0; i < 16; i++) {
1301         h->block_offset[16 + i]      =
1302         h->block_offset[32 + i]      = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 4 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1303         h->block_offset[48 + 16 + i] =
1304         h->block_offset[48 + 32 + i] = (4 * ((scan8[i] - scan8[0]) & 7) << pixel_shift) + 8 * s->uvlinesize * ((scan8[i] - scan8[0]) >> 3);
1305     }
1306
1307     /* can't be in alloc_tables because linesize isn't known there.
1308      * FIXME: redo bipred weight to not require extra buffer? */
1309     for (i = 0; i < s->slice_context_count; i++)
1310         if (h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1311             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16 * 6 * s->linesize);
1312
1313     /* Some macroblocks can be accessed before they're available in case
1314      * of lost slices, MBAFF or threading. */
1315     memset(h->slice_table, -1,
1316            (s->mb_height * s->mb_stride - 1) * sizeof(*h->slice_table));
1317
1318     // s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding ||
1319     //             s->current_picture.f.reference /* || h->contains_intra */ || 1;
1320
1321     /* We mark the current picture as non-reference after allocating it, so
1322      * that if we break out due to an error it can be released automatically
1323      * in the next ff_MPV_frame_start().
1324      * SVQ3 as well as most other codecs have only last/next/current and thus
1325      * get released even with set reference, besides SVQ3 and others do not
1326      * mark frames as reference later "naturally". */
1327     if (s->codec_id != CODEC_ID_SVQ3)
1328         s->current_picture_ptr->f.reference = 0;
1329
1330     s->current_picture_ptr->field_poc[0]     =
1331         s->current_picture_ptr->field_poc[1] = INT_MAX;
1332
1333     h->next_output_pic = NULL;
1334
1335     assert(s->current_picture_ptr->long_ref == 0);
1336
1337     return 0;
1338 }
1339
1340 /**
1341  * Run setup operations that must be run after slice header decoding.
1342  * This includes finding the next displayed frame.
1343  *
1344  * @param h h264 master context
1345  * @param setup_finished enough NALs have been read that we can call
1346  * ff_thread_finish_setup()
1347  */
1348 static void decode_postinit(H264Context *h, int setup_finished)
1349 {
1350     MpegEncContext *const s = &h->s;
1351     Picture *out = s->current_picture_ptr;
1352     Picture *cur = s->current_picture_ptr;
1353     int i, pics, out_of_order, out_idx;
1354
1355     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1356     s->current_picture_ptr->f.pict_type   = s->pict_type;
1357
1358     if (h->next_output_pic)
1359         return;
1360
1361     if (cur->field_poc[0] == INT_MAX || cur->field_poc[1] == INT_MAX) {
1362         /* FIXME: if we have two PAFF fields in one packet, we can't start
1363          * the next thread here. If we have one field per packet, we can.
1364          * The check in decode_nal_units() is not good enough to find this
1365          * yet, so we assume the worst for now. */
1366         // if (setup_finished)
1367         //    ff_thread_finish_setup(s->avctx);
1368         return;
1369     }
1370
1371     cur->f.interlaced_frame = 0;
1372     cur->f.repeat_pict      = 0;
1373
1374     /* Signal interlacing information externally. */
1375     /* Prioritize picture timing SEI information over used
1376      * decoding process if it exists. */
1377
1378     if (h->sps.pic_struct_present_flag) {
1379         switch (h->sei_pic_struct) {
1380         case SEI_PIC_STRUCT_FRAME:
1381             break;
1382         case SEI_PIC_STRUCT_TOP_FIELD:
1383         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1384             cur->f.interlaced_frame = 1;
1385             break;
1386         case SEI_PIC_STRUCT_TOP_BOTTOM:
1387         case SEI_PIC_STRUCT_BOTTOM_TOP:
1388             if (FIELD_OR_MBAFF_PICTURE)
1389                 cur->f.interlaced_frame = 1;
1390             else
1391                 // try to flag soft telecine progressive
1392                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1393             break;
1394         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1395         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1396             /* Signal the possibility of telecined film externally
1397              * (pic_struct 5,6). From these hints, let the applications
1398              * decide if they apply deinterlacing. */
1399             cur->f.repeat_pict = 1;
1400             break;
1401         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1402             // Force progressive here, doubling interlaced frame is a bad idea.
1403             cur->f.repeat_pict = 2;
1404             break;
1405         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1406             cur->f.repeat_pict = 4;
1407             break;
1408         }
1409
1410         if ((h->sei_ct_type & 3) &&
1411             h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1412             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1413     } else {
1414         /* Derive interlacing flag from used decoding process. */
1415         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1416     }
1417     h->prev_interlaced_frame = cur->f.interlaced_frame;
1418
1419     if (cur->field_poc[0] != cur->field_poc[1]) {
1420         /* Derive top_field_first from field pocs. */
1421         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1422     } else {
1423         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1424             /* Use picture timing SEI information. Even if it is a
1425              * information of a past frame, better than nothing. */
1426             if (h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM ||
1427                 h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1428                 cur->f.top_field_first = 1;
1429             else
1430                 cur->f.top_field_first = 0;
1431         } else {
1432             /* Most likely progressive */
1433             cur->f.top_field_first = 0;
1434         }
1435     }
1436
1437     cur->mmco_reset = h->mmco_reset;
1438     h->mmco_reset = 0;
1439     // FIXME do something with unavailable reference frames
1440
1441     /* Sort B-frames into display order */
1442
1443     if (h->sps.bitstream_restriction_flag &&
1444         s->avctx->has_b_frames < h->sps.num_reorder_frames) {
1445         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1446         s->low_delay           = 0;
1447     }
1448
1449     if (s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT &&
1450         !h->sps.bitstream_restriction_flag) {
1451         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1452         s->low_delay           = 0;
1453     }
1454
1455     for (i = 0; 1; i++) {
1456         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1457             if(i)
1458                 h->last_pocs[i-1] = cur->poc;
1459             break;
1460         } else if(i) {
1461             h->last_pocs[i-1]= h->last_pocs[i];
1462         }
1463     }
1464     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1465     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1466        || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
1467         out_of_order = FFMAX(out_of_order, 1);
1468     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1469         av_log(s->avctx, AV_LOG_VERBOSE, "Increasing reorder buffer to %d\n", out_of_order);
1470         s->avctx->has_b_frames = out_of_order;
1471         s->low_delay = 0;
1472     }
1473
1474     pics = 0;
1475     while (h->delayed_pic[pics])
1476         pics++;
1477
1478     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1479
1480     h->delayed_pic[pics++] = cur;
1481     if (cur->f.reference == 0)
1482         cur->f.reference = DELAYED_PIC_REF;
1483
1484     out = h->delayed_pic[0];
1485     out_idx = 0;
1486     for (i = 1; h->delayed_pic[i] &&
1487                 !h->delayed_pic[i]->f.key_frame &&
1488                 !h->delayed_pic[i]->mmco_reset;
1489          i++)
1490         if (h->delayed_pic[i]->poc < out->poc) {
1491             out     = h->delayed_pic[i];
1492             out_idx = i;
1493         }
1494     if (s->avctx->has_b_frames == 0 &&
1495         (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1496         h->next_outputed_poc = INT_MIN;
1497     out_of_order = out->poc < h->next_outputed_poc;
1498
1499     if (out_of_order || pics > s->avctx->has_b_frames) {
1500         out->f.reference &= ~DELAYED_PIC_REF;
1501         // for frame threading, the owner must be the second field's thread or
1502         // else the first thread can release the picture and reuse it unsafely
1503         out->owner2       = s;
1504         for (i = out_idx; h->delayed_pic[i]; i++)
1505             h->delayed_pic[i] = h->delayed_pic[i + 1];
1506     }
1507     if (!out_of_order && pics > s->avctx->has_b_frames) {
1508         h->next_output_pic = out;
1509         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1510             h->next_outputed_poc = INT_MIN;
1511         } else
1512             h->next_outputed_poc = out->poc;
1513     } else {
1514         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1515     }
1516
1517     if (h->next_output_pic && h->next_output_pic->sync) {
1518         h->sync |= 2;
1519     }
1520
1521     if (setup_finished)
1522         ff_thread_finish_setup(s->avctx);
1523 }
1524
1525 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1526                                               uint8_t *src_cb, uint8_t *src_cr,
1527                                               int linesize, int uvlinesize,
1528                                               int simple)
1529 {
1530     MpegEncContext *const s = &h->s;
1531     uint8_t *top_border;
1532     int top_idx = 1;
1533     const int pixel_shift = h->pixel_shift;
1534     int chroma444 = CHROMA444;
1535     int chroma422 = CHROMA422;
1536
1537     src_y  -= linesize;
1538     src_cb -= uvlinesize;
1539     src_cr -= uvlinesize;
1540
1541     if (!simple && FRAME_MBAFF) {
1542         if (s->mb_y & 1) {
1543             if (!MB_MBAFF) {
1544                 top_border = h->top_borders[0][s->mb_x];
1545                 AV_COPY128(top_border, src_y + 15 * linesize);
1546                 if (pixel_shift)
1547                     AV_COPY128(top_border + 16, src_y + 15 * linesize + 16);
1548                 if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1549                     if (chroma444) {
1550                         if (pixel_shift) {
1551                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1552                             AV_COPY128(top_border + 48, src_cb + 15 * uvlinesize + 16);
1553                             AV_COPY128(top_border + 64, src_cr + 15 * uvlinesize);
1554                             AV_COPY128(top_border + 80, src_cr + 15 * uvlinesize + 16);
1555                         } else {
1556                             AV_COPY128(top_border + 16, src_cb + 15 * uvlinesize);
1557                             AV_COPY128(top_border + 32, src_cr + 15 * uvlinesize);
1558                         }
1559                     } else if (chroma422) {
1560                         if (pixel_shift) {
1561                             AV_COPY128(top_border + 32, src_cb + 15 * uvlinesize);
1562                             AV_COPY128(top_border + 48, src_cr + 15 * uvlinesize);
1563                         } else {
1564                             AV_COPY64(top_border + 16, src_cb + 15 * uvlinesize);
1565                             AV_COPY64(top_border + 24, src_cr + 15 * uvlinesize);
1566                         }
1567                     } else {
1568                         if (pixel_shift) {
1569                             AV_COPY128(top_border + 32, src_cb + 7 * uvlinesize);
1570                             AV_COPY128(top_border + 48, src_cr + 7 * uvlinesize);
1571                         } else {
1572                             AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
1573                             AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
1574                         }
1575                     }
1576                 }
1577             }
1578         } else if (MB_MBAFF) {
1579             top_idx = 0;
1580         } else
1581             return;
1582     }
1583
1584     top_border = h->top_borders[top_idx][s->mb_x];
1585     /* There are two lines saved, the line above the top macroblock
1586      * of a pair, and the line above the bottom macroblock. */
1587     AV_COPY128(top_border, src_y + 16 * linesize);
1588     if (pixel_shift)
1589         AV_COPY128(top_border + 16, src_y + 16 * linesize + 16);
1590
1591     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1592         if (chroma444) {
1593             if (pixel_shift) {
1594                 AV_COPY128(top_border + 32, src_cb + 16 * linesize);
1595                 AV_COPY128(top_border + 48, src_cb + 16 * linesize + 16);
1596                 AV_COPY128(top_border + 64, src_cr + 16 * linesize);
1597                 AV_COPY128(top_border + 80, src_cr + 16 * linesize + 16);
1598             } else {
1599                 AV_COPY128(top_border + 16, src_cb + 16 * linesize);
1600                 AV_COPY128(top_border + 32, src_cr + 16 * linesize);
1601             }
1602         } else if (chroma422) {
1603             if (pixel_shift) {
1604                 AV_COPY128(top_border + 32, src_cb + 16 * uvlinesize);
1605                 AV_COPY128(top_border + 48, src_cr + 16 * uvlinesize);
1606             } else {
1607                 AV_COPY64(top_border + 16, src_cb + 16 * uvlinesize);
1608                 AV_COPY64(top_border + 24, src_cr + 16 * uvlinesize);
1609             }
1610         } else {
1611             if (pixel_shift) {
1612                 AV_COPY128(top_border + 32, src_cb + 8 * uvlinesize);
1613                 AV_COPY128(top_border + 48, src_cr + 8 * uvlinesize);
1614             } else {
1615                 AV_COPY64(top_border + 16, src_cb + 8 * uvlinesize);
1616                 AV_COPY64(top_border + 24, src_cr + 8 * uvlinesize);
1617             }
1618         }
1619     }
1620 }
1621
1622 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1623                                             uint8_t *src_cb, uint8_t *src_cr,
1624                                             int linesize, int uvlinesize,
1625                                             int xchg, int chroma444,
1626                                             int simple, int pixel_shift)
1627 {
1628     MpegEncContext *const s = &h->s;
1629     int deblock_topleft;
1630     int deblock_top;
1631     int top_idx = 1;
1632     uint8_t *top_border_m1;
1633     uint8_t *top_border;
1634
1635     if (!simple && FRAME_MBAFF) {
1636         if (s->mb_y & 1) {
1637             if (!MB_MBAFF)
1638                 return;
1639         } else {
1640             top_idx = MB_MBAFF ? 0 : 1;
1641         }
1642     }
1643
1644     if (h->deblocking_filter == 2) {
1645         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1646         deblock_top     = h->top_type;
1647     } else {
1648         deblock_topleft = (s->mb_x > 0);
1649         deblock_top     = (s->mb_y > !!MB_FIELD);
1650     }
1651
1652     src_y  -= linesize   + 1 + pixel_shift;
1653     src_cb -= uvlinesize + 1 + pixel_shift;
1654     src_cr -= uvlinesize + 1 + pixel_shift;
1655
1656     top_border_m1 = h->top_borders[top_idx][s->mb_x - 1];
1657     top_border    = h->top_borders[top_idx][s->mb_x];
1658
1659 #define XCHG(a, b, xchg)                        \
1660     if (pixel_shift) {                          \
1661         if (xchg) {                             \
1662             AV_SWAP64(b + 0, a + 0);            \
1663             AV_SWAP64(b + 8, a + 8);            \
1664         } else {                                \
1665             AV_COPY128(b, a);                   \
1666         }                                       \
1667     } else if (xchg)                            \
1668         AV_SWAP64(b, a);                        \
1669     else                                        \
1670         AV_COPY64(b, a);
1671
1672     if (deblock_top) {
1673         if (deblock_topleft) {
1674             XCHG(top_border_m1 + (8 << pixel_shift),
1675                  src_y - (7 << pixel_shift), 1);
1676         }
1677         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1678         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1679         if (s->mb_x + 1 < s->mb_width) {
1680             XCHG(h->top_borders[top_idx][s->mb_x + 1],
1681                  src_y + (17 << pixel_shift), 1);
1682         }
1683     }
1684     if (simple || !CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) {
1685         if (chroma444) {
1686             if (deblock_topleft) {
1687                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1688                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1689             }
1690             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1691             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1692             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1693             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1694             if (s->mb_x + 1 < s->mb_width) {
1695                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1696                 XCHG(h->top_borders[top_idx][s->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1697             }
1698         } else {
1699             if (deblock_top) {
1700                 if (deblock_topleft) {
1701                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1702                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1703                 }
1704                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
1705                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
1706             }
1707         }
1708     }
1709 }
1710
1711 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth,
1712                                         int index)
1713 {
1714     if (high_bit_depth) {
1715         return AV_RN32A(((int32_t *)mb) + index);
1716     } else
1717         return AV_RN16A(mb + index);
1718 }
1719
1720 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth,
1721                                          int index, int value)
1722 {
1723     if (high_bit_depth) {
1724         AV_WN32A(((int32_t *)mb) + index, value);
1725     } else
1726         AV_WN16A(mb + index, value);
1727 }
1728
1729 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
1730                                                        int mb_type, int is_h264,
1731                                                        int simple,
1732                                                        int transform_bypass,
1733                                                        int pixel_shift,
1734                                                        int *block_offset,
1735                                                        int linesize,
1736                                                        uint8_t *dest_y, int p)
1737 {
1738     MpegEncContext *const s = &h->s;
1739     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1740     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1741     int i;
1742     int qscale = p == 0 ? s->qscale : h->chroma_qp[p - 1];
1743     block_offset += 16 * p;
1744     if (IS_INTRA4x4(mb_type)) {
1745         if (simple || !s->encoding) {
1746             if (IS_8x8DCT(mb_type)) {
1747                 if (transform_bypass) {
1748                     idct_dc_add  =
1749                     idct_add     = s->dsp.add_pixels8;
1750                 } else {
1751                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1752                     idct_add    = h->h264dsp.h264_idct8_add;
1753                 }
1754                 for (i = 0; i < 16; i += 4) {
1755                     uint8_t *const ptr = dest_y + block_offset[i];
1756                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1757                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1758                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1759                     } else {
1760                         const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1761                         h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
1762                                              (h->topright_samples_available << i) & 0x4000, linesize);
1763                         if (nnz) {
1764                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1765                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1766                             else
1767                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1768                         }
1769                     }
1770                 }
1771             } else {
1772                 if (transform_bypass) {
1773                     idct_dc_add  =
1774                         idct_add = s->dsp.add_pixels4;
1775                 } else {
1776                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1777                     idct_add    = h->h264dsp.h264_idct_add;
1778                 }
1779                 for (i = 0; i < 16; i++) {
1780                     uint8_t *const ptr = dest_y + block_offset[i];
1781                     const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
1782
1783                     if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
1784                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1785                     } else {
1786                         uint8_t *topright;
1787                         int nnz, tr;
1788                         uint64_t tr_high;
1789                         if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
1790                             const int topright_avail = (h->topright_samples_available << i) & 0x8000;
1791                             assert(s->mb_y || linesize <= block_offset[i]);
1792                             if (!topright_avail) {
1793                                 if (pixel_shift) {
1794                                     tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
1795                                     topright = (uint8_t *)&tr_high;
1796                                 } else {
1797                                     tr       = ptr[3 - linesize] * 0x01010101u;
1798                                     topright = (uint8_t *)&tr;
1799                                 }
1800                             } else
1801                                 topright = ptr + (4 << pixel_shift) - linesize;
1802                         } else
1803                             topright = NULL;
1804
1805                         h->hpc.pred4x4[dir](ptr, topright, linesize);
1806                         nnz = h->non_zero_count_cache[scan8[i + p * 16]];
1807                         if (nnz) {
1808                             if (is_h264) {
1809                                 if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1810                                     idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1811                                 else
1812                                     idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
1813                             } else if (CONFIG_SVQ3_DECODER)
1814                                 ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
1815                         }
1816                     }
1817                 }
1818             }
1819         }
1820     } else {
1821         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
1822         if (is_h264) {
1823             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
1824                 if (!transform_bypass)
1825                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
1826                                                          h->mb_luma_dc[p],
1827                                                          h->dequant4_coeff[p][qscale][0]);
1828                 else {
1829                     static const uint8_t dc_mapping[16] = {
1830                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
1831                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
1832                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
1833                         10 * 16, 11 * 16, 14 * 16, 15 * 16 };
1834                     for (i = 0; i < 16; i++)
1835                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
1836                                     pixel_shift, dc_mapping[i],
1837                                     dctcoef_get(h->mb_luma_dc[p],
1838                                                 pixel_shift, i));
1839                 }
1840             }
1841         } else if (CONFIG_SVQ3_DECODER)
1842             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
1843                                            h->mb_luma_dc[p], qscale);
1844     }
1845 }
1846
1847 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
1848                                                     int is_h264, int simple,
1849                                                     int transform_bypass,
1850                                                     int pixel_shift,
1851                                                     int *block_offset,
1852                                                     int linesize,
1853                                                     uint8_t *dest_y, int p)
1854 {
1855     MpegEncContext *const s = &h->s;
1856     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1857     int i;
1858     block_offset += 16 * p;
1859     if (!IS_INTRA4x4(mb_type)) {
1860         if (is_h264) {
1861             if (IS_INTRA16x16(mb_type)) {
1862                 if (transform_bypass) {
1863                     if (h->sps.profile_idc == 244 &&
1864                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
1865                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
1866                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
1867                                                                       h->mb + (p * 256 << pixel_shift),
1868                                                                       linesize);
1869                     } else {
1870                         for (i = 0; i < 16; i++)
1871                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
1872                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
1873                                 s->dsp.add_pixels4(dest_y + block_offset[i],
1874                                                    h->mb + (i * 16 + p * 256 << pixel_shift),
1875                                                    linesize);
1876                     }
1877                 } else {
1878                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
1879                                                     h->mb + (p * 256 << pixel_shift),
1880                                                     linesize,
1881                                                     h->non_zero_count_cache + p * 5 * 8);
1882                 }
1883             } else if (h->cbp & 15) {
1884                 if (transform_bypass) {
1885                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1886                     idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8
1887                                                   : s->dsp.add_pixels4;
1888                     for (i = 0; i < 16; i += di)
1889                         if (h->non_zero_count_cache[scan8[i + p * 16]])
1890                             idct_add(dest_y + block_offset[i],
1891                                      h->mb + (i * 16 + p * 256 << pixel_shift),
1892                                      linesize);
1893                 } else {
1894                     if (IS_8x8DCT(mb_type))
1895                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
1896                                                    h->mb + (p * 256 << pixel_shift),
1897                                                    linesize,
1898                                                    h->non_zero_count_cache + p * 5 * 8);
1899                     else
1900                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
1901                                                    h->mb + (p * 256 << pixel_shift),
1902                                                    linesize,
1903                                                    h->non_zero_count_cache + p * 5 * 8);
1904                 }
1905             }
1906         } else if (CONFIG_SVQ3_DECODER) {
1907             for (i = 0; i < 16; i++)
1908                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
1909                     // FIXME benchmark weird rule, & below
1910                     uint8_t *const ptr = dest_y + block_offset[i];
1911                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
1912                                        s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1913                 }
1914         }
1915     }
1916 }
1917
1918 #define BITS   8
1919 #define SIMPLE 1
1920 #include "h264_mb_template.c"
1921
1922 #undef  BITS
1923 #define BITS   16
1924 #include "h264_mb_template.c"
1925
1926 #undef  SIMPLE
1927 #define SIMPLE 0
1928 #include "h264_mb_template.c"
1929
1930 void ff_h264_hl_decode_mb(H264Context *h)
1931 {
1932     MpegEncContext *const s = &h->s;
1933     const int mb_xy   = h->mb_xy;
1934     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1935     int is_complex    = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1936
1937     if (CHROMA444) {
1938         if (is_complex || h->pixel_shift)
1939             hl_decode_mb_444_complex(h);
1940         else
1941             hl_decode_mb_444_simple_8(h);
1942     } else if (is_complex) {
1943         hl_decode_mb_complex(h);
1944     } else if (h->pixel_shift) {
1945         hl_decode_mb_simple_16(h);
1946     } else
1947         hl_decode_mb_simple_8(h);
1948 }
1949
1950 static int pred_weight_table(H264Context *h)
1951 {
1952     MpegEncContext *const s = &h->s;
1953     int list, i;
1954     int luma_def, chroma_def;
1955
1956     h->use_weight             = 0;
1957     h->use_weight_chroma      = 0;
1958     h->luma_log2_weight_denom = get_ue_golomb(&s->gb);
1959     if (h->sps.chroma_format_idc)
1960         h->chroma_log2_weight_denom = get_ue_golomb(&s->gb);
1961     luma_def   = 1 << h->luma_log2_weight_denom;
1962     chroma_def = 1 << h->chroma_log2_weight_denom;
1963
1964     for (list = 0; list < 2; list++) {
1965         h->luma_weight_flag[list]   = 0;
1966         h->chroma_weight_flag[list] = 0;
1967         for (i = 0; i < h->ref_count[list]; i++) {
1968             int luma_weight_flag, chroma_weight_flag;
1969
1970             luma_weight_flag = get_bits1(&s->gb);
1971             if (luma_weight_flag) {
1972                 h->luma_weight[i][list][0] = get_se_golomb(&s->gb);
1973                 h->luma_weight[i][list][1] = get_se_golomb(&s->gb);
1974                 if (h->luma_weight[i][list][0] != luma_def ||
1975                     h->luma_weight[i][list][1] != 0) {
1976                     h->use_weight             = 1;
1977                     h->luma_weight_flag[list] = 1;
1978                 }
1979             } else {
1980                 h->luma_weight[i][list][0] = luma_def;
1981                 h->luma_weight[i][list][1] = 0;
1982             }
1983
1984             if (h->sps.chroma_format_idc) {
1985                 chroma_weight_flag = get_bits1(&s->gb);
1986                 if (chroma_weight_flag) {
1987                     int j;
1988                     for (j = 0; j < 2; j++) {
1989                         h->chroma_weight[i][list][j][0] = get_se_golomb(&s->gb);
1990                         h->chroma_weight[i][list][j][1] = get_se_golomb(&s->gb);
1991                         if (h->chroma_weight[i][list][j][0] != chroma_def ||
1992                             h->chroma_weight[i][list][j][1] != 0) {
1993                             h->use_weight_chroma = 1;
1994                             h->chroma_weight_flag[list] = 1;
1995                         }
1996                     }
1997                 } else {
1998                     int j;
1999                     for (j = 0; j < 2; j++) {
2000                         h->chroma_weight[i][list][j][0] = chroma_def;
2001                         h->chroma_weight[i][list][j][1] = 0;
2002                     }
2003                 }
2004             }
2005         }
2006         if (h->slice_type_nos != AV_PICTURE_TYPE_B)
2007             break;
2008     }
2009     h->use_weight = h->use_weight || h->use_weight_chroma;
2010     return 0;
2011 }
2012
2013 /**
2014  * Initialize implicit_weight table.
2015  * @param field  0/1 initialize the weight for interlaced MBAFF
2016  *                -1 initializes the rest
2017  */
2018 static void implicit_weight_table(H264Context *h, int field)
2019 {
2020     MpegEncContext *const s = &h->s;
2021     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2022
2023     for (i = 0; i < 2; i++) {
2024         h->luma_weight_flag[i]   = 0;
2025         h->chroma_weight_flag[i] = 0;
2026     }
2027
2028     if (field < 0) {
2029         if (s->picture_structure == PICT_FRAME) {
2030             cur_poc = s->current_picture_ptr->poc;
2031         } else {
2032             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2033         }
2034         if (h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF &&
2035             h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2 * cur_poc) {
2036             h->use_weight = 0;
2037             h->use_weight_chroma = 0;
2038             return;
2039         }
2040         ref_start  = 0;
2041         ref_count0 = h->ref_count[0];
2042         ref_count1 = h->ref_count[1];
2043     } else {
2044         cur_poc    = s->current_picture_ptr->field_poc[field];
2045         ref_start  = 16;
2046         ref_count0 = 16 + 2 * h->ref_count[0];
2047         ref_count1 = 16 + 2 * h->ref_count[1];
2048     }
2049
2050     h->use_weight               = 2;
2051     h->use_weight_chroma        = 2;
2052     h->luma_log2_weight_denom   = 5;
2053     h->chroma_log2_weight_denom = 5;
2054
2055     for (ref0 = ref_start; ref0 < ref_count0; ref0++) {
2056         int poc0 = h->ref_list[0][ref0].poc;
2057         for (ref1 = ref_start; ref1 < ref_count1; ref1++) {
2058             int w = 32;
2059             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2060                 int poc1 = h->ref_list[1][ref1].poc;
2061                 int td   = av_clip(poc1 - poc0, -128, 127);
2062                 if (td) {
2063                     int tb = av_clip(cur_poc - poc0, -128, 127);
2064                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2065                     int dist_scale_factor = (tb * tx + 32) >> 8;
2066                     if (dist_scale_factor >= -64 && dist_scale_factor <= 128)
2067                         w = 64 - dist_scale_factor;
2068                 }
2069             }
2070             if (field < 0) {
2071                 h->implicit_weight[ref0][ref1][0] =
2072                 h->implicit_weight[ref0][ref1][1] = w;
2073             } else {
2074                 h->implicit_weight[ref0][ref1][field] = w;
2075             }
2076         }
2077     }
2078 }
2079
2080 /**
2081  * instantaneous decoder refresh.
2082  */
2083 static void idr(H264Context *h)
2084 {
2085     int i;
2086     ff_h264_remove_all_refs(h);
2087     h->prev_frame_num        = 0;
2088     h->prev_frame_num_offset = 0;
2089     h->prev_poc_msb          = 1<<16;
2090     h->prev_poc_lsb          = 0;
2091     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2092         h->last_pocs[i] = INT_MIN;
2093 }
2094
2095 /* forget old pics after a seek */
2096 static void flush_dpb(AVCodecContext *avctx)
2097 {
2098     H264Context *h = avctx->priv_data;
2099     int i;
2100     for (i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2101         if (h->delayed_pic[i])
2102             h->delayed_pic[i]->f.reference = 0;
2103         h->delayed_pic[i] = NULL;
2104     }
2105     h->outputed_poc = h->next_outputed_poc = INT_MIN;
2106     h->prev_interlaced_frame = 1;
2107     idr(h);
2108     h->prev_frame_num = -1;
2109     if (h->s.current_picture_ptr)
2110         h->s.current_picture_ptr->f.reference = 0;
2111     h->s.first_field = 0;
2112     ff_h264_reset_sei(h);
2113     ff_mpeg_flush(avctx);
2114     h->recovery_frame= -1;
2115     h->sync= 0;
2116 }
2117
2118 static int init_poc(H264Context *h)
2119 {
2120     MpegEncContext *const s = &h->s;
2121     const int max_frame_num = 1 << h->sps.log2_max_frame_num;
2122     int field_poc[2];
2123     Picture *cur = s->current_picture_ptr;
2124
2125     h->frame_num_offset = h->prev_frame_num_offset;
2126     if (h->frame_num < h->prev_frame_num)
2127         h->frame_num_offset += max_frame_num;
2128
2129     if (h->sps.poc_type == 0) {
2130         const int max_poc_lsb = 1 << h->sps.log2_max_poc_lsb;
2131
2132         if (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb / 2)
2133             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2134         else if (h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb / 2)
2135             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2136         else
2137             h->poc_msb = h->prev_poc_msb;
2138         // printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2139         field_poc[0] =
2140         field_poc[1] = h->poc_msb + h->poc_lsb;
2141         if (s->picture_structure == PICT_FRAME)
2142             field_poc[1] += h->delta_poc_bottom;
2143     } else if (h->sps.poc_type == 1) {
2144         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2145         int i;
2146
2147         if (h->sps.poc_cycle_length != 0)
2148             abs_frame_num = h->frame_num_offset + h->frame_num;
2149         else
2150             abs_frame_num = 0;
2151
2152         if (h->nal_ref_idc == 0 && abs_frame_num > 0)
2153             abs_frame_num--;
2154
2155         expected_delta_per_poc_cycle = 0;
2156         for (i = 0; i < h->sps.poc_cycle_length; i++)
2157             // FIXME integrate during sps parse
2158             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[i];
2159
2160         if (abs_frame_num > 0) {
2161             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2162             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2163
2164             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2165             for (i = 0; i <= frame_num_in_poc_cycle; i++)
2166                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[i];
2167         } else
2168             expectedpoc = 0;
2169
2170         if (h->nal_ref_idc == 0)
2171             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2172
2173         field_poc[0] = expectedpoc + h->delta_poc[0];
2174         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2175
2176         if (s->picture_structure == PICT_FRAME)
2177             field_poc[1] += h->delta_poc[1];
2178     } else {
2179         int poc = 2 * (h->frame_num_offset + h->frame_num);
2180
2181         if (!h->nal_ref_idc)
2182             poc--;
2183
2184         field_poc[0] = poc;
2185         field_poc[1] = poc;
2186     }
2187
2188     if (s->picture_structure != PICT_BOTTOM_FIELD)
2189         s->current_picture_ptr->field_poc[0] = field_poc[0];
2190     if (s->picture_structure != PICT_TOP_FIELD)
2191         s->current_picture_ptr->field_poc[1] = field_poc[1];
2192     cur->poc = FFMIN(cur->field_poc[0], cur->field_poc[1]);
2193
2194     return 0;
2195 }
2196
2197 /**
2198  * initialize scan tables
2199  */
2200 static void init_scan_tables(H264Context *h)
2201 {
2202     int i;
2203     for (i = 0; i < 16; i++) {
2204 #define T(x) (x >> 2) | ((x << 2) & 0xF)
2205         h->zigzag_scan[i] = T(zigzag_scan[i]);
2206         h->field_scan[i]  = T(field_scan[i]);
2207 #undef T
2208     }
2209     for (i = 0; i < 64; i++) {
2210 #define T(x) (x >> 3) | ((x & 7) << 3)
2211         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2212         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2213         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2214         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2215 #undef T
2216     }
2217     if (h->sps.transform_bypass) { // FIXME same ugly
2218         memcpy(h->zigzag_scan_q0          , zigzag_scan             , sizeof(h->zigzag_scan_q0         ));
2219         memcpy(h->zigzag_scan8x8_q0       , ff_zigzag_direct        , sizeof(h->zigzag_scan8x8_q0      ));
2220         memcpy(h->zigzag_scan8x8_cavlc_q0 , zigzag_scan8x8_cavlc    , sizeof(h->zigzag_scan8x8_cavlc_q0));
2221         memcpy(h->field_scan_q0           , field_scan              , sizeof(h->field_scan_q0          ));
2222         memcpy(h->field_scan8x8_q0        , field_scan8x8           , sizeof(h->field_scan8x8_q0       ));
2223         memcpy(h->field_scan8x8_cavlc_q0  , field_scan8x8_cavlc     , sizeof(h->field_scan8x8_cavlc_q0 ));
2224     } else {
2225         memcpy(h->zigzag_scan_q0          , h->zigzag_scan          , sizeof(h->zigzag_scan_q0         ));
2226         memcpy(h->zigzag_scan8x8_q0       , h->zigzag_scan8x8       , sizeof(h->zigzag_scan8x8_q0      ));
2227         memcpy(h->zigzag_scan8x8_cavlc_q0 , h->zigzag_scan8x8_cavlc , sizeof(h->zigzag_scan8x8_cavlc_q0));
2228         memcpy(h->field_scan_q0           , h->field_scan           , sizeof(h->field_scan_q0          ));
2229         memcpy(h->field_scan8x8_q0        , h->field_scan8x8        , sizeof(h->field_scan8x8_q0       ));
2230         memcpy(h->field_scan8x8_cavlc_q0  , h->field_scan8x8_cavlc  , sizeof(h->field_scan8x8_cavlc_q0 ));
2231     }
2232 }
2233
2234 static int field_end(H264Context *h, int in_setup)
2235 {
2236     MpegEncContext *const s     = &h->s;
2237     AVCodecContext *const avctx = s->avctx;
2238     int err = 0;
2239     s->mb_y = 0;
2240
2241     if (!in_setup && !s->dropable)
2242         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2243                                   s->picture_structure == PICT_BOTTOM_FIELD);
2244
2245     if (CONFIG_H264_VDPAU_DECODER &&
2246         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2247         ff_vdpau_h264_set_reference_frames(s);
2248
2249     if (in_setup || !(avctx->active_thread_type & FF_THREAD_FRAME)) {
2250         if (!s->dropable) {
2251             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2252             h->prev_poc_msb = h->poc_msb;
2253             h->prev_poc_lsb = h->poc_lsb;
2254         }
2255         h->prev_frame_num_offset = h->frame_num_offset;
2256         h->prev_frame_num        = h->frame_num;
2257         h->outputed_poc          = h->next_outputed_poc;
2258     }
2259
2260     if (avctx->hwaccel) {
2261         if (avctx->hwaccel->end_frame(avctx) < 0)
2262             av_log(avctx, AV_LOG_ERROR,
2263                    "hardware accelerator failed to decode picture\n");
2264     }
2265
2266     if (CONFIG_H264_VDPAU_DECODER &&
2267         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
2268         ff_vdpau_h264_picture_complete(s);
2269
2270     /*
2271      * FIXME: Error handling code does not seem to support interlaced
2272      * when slices span multiple rows
2273      * The ff_er_add_slice calls don't work right for bottom
2274      * fields; they cause massive erroneous error concealing
2275      * Error marking covers both fields (top and bottom).
2276      * This causes a mismatched s->error_count
2277      * and a bad error table. Further, the error count goes to
2278      * INT_MAX when called for bottom field, because mb_y is
2279      * past end by one (callers fault) and resync_mb_y != 0
2280      * causes problems for the first MB line, too.
2281      */
2282     if (!FIELD_PICTURE)
2283         ff_er_frame_end(s);
2284
2285     ff_MPV_frame_end(s);
2286
2287     h->current_slice = 0;
2288
2289     return err;
2290 }
2291
2292 /**
2293  * Replicate H264 "master" context to thread contexts.
2294  */
2295 static void clone_slice(H264Context *dst, H264Context *src)
2296 {
2297     memcpy(dst->block_offset, src->block_offset, sizeof(dst->block_offset));
2298     dst->s.current_picture_ptr = src->s.current_picture_ptr;
2299     dst->s.current_picture     = src->s.current_picture;
2300     dst->s.linesize            = src->s.linesize;
2301     dst->s.uvlinesize          = src->s.uvlinesize;
2302     dst->s.first_field         = src->s.first_field;
2303
2304     dst->prev_poc_msb          = src->prev_poc_msb;
2305     dst->prev_poc_lsb          = src->prev_poc_lsb;
2306     dst->prev_frame_num_offset = src->prev_frame_num_offset;
2307     dst->prev_frame_num        = src->prev_frame_num;
2308     dst->short_ref_count       = src->short_ref_count;
2309
2310     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2311     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2312     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2313     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2314
2315     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2316     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2317 }
2318
2319 /**
2320  * Compute profile from profile_idc and constraint_set?_flags.
2321  *
2322  * @param sps SPS
2323  *
2324  * @return profile as defined by FF_PROFILE_H264_*
2325  */
2326 int ff_h264_get_profile(SPS *sps)
2327 {
2328     int profile = sps->profile_idc;
2329
2330     switch (sps->profile_idc) {
2331     case FF_PROFILE_H264_BASELINE:
2332         // constraint_set1_flag set to 1
2333         profile |= (sps->constraint_set_flags & 1 << 1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2334         break;
2335     case FF_PROFILE_H264_HIGH_10:
2336     case FF_PROFILE_H264_HIGH_422:
2337     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2338         // constraint_set3_flag set to 1
2339         profile |= (sps->constraint_set_flags & 1 << 3) ? FF_PROFILE_H264_INTRA : 0;
2340         break;
2341     }
2342
2343     return profile;
2344 }
2345
2346 /**
2347  * Decode a slice header.
2348  * This will also call ff_MPV_common_init() and frame_start() as needed.
2349  *
2350  * @param h h264context
2351  * @param h0 h264 master context (differs from 'h' when doing sliced based
2352  *           parallel decoding)
2353  *
2354  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2355  */
2356 static int decode_slice_header(H264Context *h, H264Context *h0)
2357 {
2358     MpegEncContext *const s  = &h->s;
2359     MpegEncContext *const s0 = &h0->s;
2360     unsigned int first_mb_in_slice;
2361     unsigned int pps_id;
2362     int num_ref_idx_active_override_flag;
2363     unsigned int slice_type, tmp, i, j;
2364     int default_ref_list_done = 0;
2365     int last_pic_structure, last_pic_dropable;
2366     int must_reinit;
2367
2368     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2369     if ((s->avctx->flags2 & CODEC_FLAG2_FAST) &&
2370         !h->nal_ref_idc && !h->pixel_shift) {
2371         s->me.qpel_put = s->dsp.put_2tap_qpel_pixels_tab;
2372         s->me.qpel_avg = s->dsp.avg_2tap_qpel_pixels_tab;
2373     } else {
2374         s->me.qpel_put = s->dsp.put_h264_qpel_pixels_tab;
2375         s->me.qpel_avg = s->dsp.avg_h264_qpel_pixels_tab;
2376     }
2377
2378     first_mb_in_slice = get_ue_golomb_long(&s->gb);
2379
2380     if (first_mb_in_slice == 0) { // FIXME better field boundary detection
2381         if (h0->current_slice && FIELD_PICTURE) {
2382             field_end(h, 1);
2383         }
2384
2385         h0->current_slice = 0;
2386         if (!s0->first_field) {
2387             if (s->current_picture_ptr && !s->dropable &&
2388                 s->current_picture_ptr->owner2 == s) {
2389                 ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
2390                                           s->picture_structure == PICT_BOTTOM_FIELD);
2391             }
2392             s->current_picture_ptr = NULL;
2393         }
2394     }
2395
2396     slice_type = get_ue_golomb_31(&s->gb);
2397     if (slice_type > 9) {
2398         av_log(h->s.avctx, AV_LOG_ERROR,
2399                "slice type too large (%d) at %d %d\n",
2400                h->slice_type, s->mb_x, s->mb_y);
2401         return -1;
2402     }
2403     if (slice_type > 4) {
2404         slice_type -= 5;
2405         h->slice_type_fixed = 1;
2406     } else
2407         h->slice_type_fixed = 0;
2408
2409     slice_type = golomb_to_pict_type[slice_type];
2410     if (slice_type == AV_PICTURE_TYPE_I ||
2411         (h0->current_slice != 0 && slice_type == h0->last_slice_type)) {
2412         default_ref_list_done = 1;
2413     }
2414     h->slice_type     = slice_type;
2415     h->slice_type_nos = slice_type & 3;
2416
2417     // to make a few old functions happy, it's wrong though
2418     s->pict_type = h->slice_type;
2419
2420     pps_id = get_ue_golomb(&s->gb);
2421     if (pps_id >= MAX_PPS_COUNT) {
2422         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id %d out of range\n", pps_id);
2423         return -1;
2424     }
2425     if (!h0->pps_buffers[pps_id]) {
2426         av_log(h->s.avctx, AV_LOG_ERROR,
2427                "non-existing PPS %u referenced\n",
2428                pps_id);
2429         return -1;
2430     }
2431     h->pps = *h0->pps_buffers[pps_id];
2432
2433     if (!h0->sps_buffers[h->pps.sps_id]) {
2434         av_log(h->s.avctx, AV_LOG_ERROR,
2435                "non-existing SPS %u referenced\n",
2436                h->pps.sps_id);
2437         return -1;
2438     }
2439     h->sps = *h0->sps_buffers[h->pps.sps_id];
2440
2441     s->avctx->profile = ff_h264_get_profile(&h->sps);
2442     s->avctx->level   = h->sps.level_idc;
2443     s->avctx->refs    = h->sps.ref_frame_count;
2444
2445     must_reinit = (s->context_initialized &&
2446                     (   16*h->sps.mb_width != s->avctx->coded_width
2447                      || 16*h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag) != s->avctx->coded_height
2448                      || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2449                      || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2450                      || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio)));
2451
2452     if(must_reinit && (h != h0 || (s->avctx->active_thread_type & FF_THREAD_FRAME))) {
2453         av_log_missing_feature(s->avctx,
2454                                 "Width/height/bit depth/chroma idc changing with threads is", 0);
2455         return AVERROR_PATCHWELCOME;   // width / height changed during parallelized decoding
2456     }
2457
2458     s->mb_width  = h->sps.mb_width;
2459     s->mb_height = h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2460
2461     h->b_stride = s->mb_width * 4;
2462
2463     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2464
2465     s->width  = 16 * s->mb_width;
2466     s->height = 16 * s->mb_height;
2467
2468     if(must_reinit) {
2469         free_tables(h, 0);
2470         flush_dpb(s->avctx);
2471         ff_MPV_common_end(s);
2472         h->list_count = 0;
2473         h->current_slice = 0;
2474     }
2475     if (!s->context_initialized) {
2476         if (h != h0) {
2477             av_log(h->s.avctx, AV_LOG_ERROR,
2478                    "Cannot (re-)initialize context during parallel decoding.\n");
2479             return -1;
2480         }
2481         if(   FFALIGN(s->avctx->width , 16                                 ) == s->width
2482            && FFALIGN(s->avctx->height, 16*(2 - h->sps.frame_mbs_only_flag)) == s->height
2483            && !h->sps.crop_right && !h->sps.crop_bottom
2484            && (s->avctx->width != s->width || s->avctx->height && s->height)
2485         ) {
2486             av_log(h->s.avctx, AV_LOG_DEBUG, "Using externally provided dimensions\n");
2487             s->avctx->coded_width  = s->width;
2488             s->avctx->coded_height = s->height;
2489         } else{
2490             avcodec_set_dimensions(s->avctx, s->width, s->height);
2491             s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2492             s->avctx->height -= (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1) * (2 - h->sps.frame_mbs_only_flag);
2493         }
2494         s->avctx->sample_aspect_ratio = h->sps.sar;
2495         av_assert0(s->avctx->sample_aspect_ratio.den);
2496
2497         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2498             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2499             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 14 && h->sps.bit_depth_luma != 11 && h->sps.bit_depth_luma != 13 &&
2500                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2501                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2502                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2503                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2504
2505                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2506                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2507                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2508                 ff_dsputil_init(&s->dsp, s->avctx);
2509             } else {
2510                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
2511                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2512                 return -1;
2513             }
2514         }
2515
2516         if (h->sps.video_signal_type_present_flag) {
2517             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG
2518                                                       : AVCOL_RANGE_MPEG;
2519             if (h->sps.colour_description_present_flag) {
2520                 s->avctx->color_primaries = h->sps.color_primaries;
2521                 s->avctx->color_trc       = h->sps.color_trc;
2522                 s->avctx->colorspace      = h->sps.colorspace;
2523             }
2524         }
2525
2526         if (h->sps.timing_info_present_flag) {
2527             int64_t den = h->sps.time_scale;
2528             if (h->x264_build < 44U)
2529                 den *= 2;
2530             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2531                       h->sps.num_units_in_tick, den, 1 << 30);
2532         }
2533
2534         switch (h->sps.bit_depth_luma) {
2535         case 9:
2536             if (CHROMA444) {
2537                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2538                     s->avctx->pix_fmt = PIX_FMT_GBRP9;
2539                 } else
2540                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2541             } else if (CHROMA422)
2542                 s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2543             else
2544                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2545             break;
2546         case 10:
2547             if (CHROMA444) {
2548                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2549                     s->avctx->pix_fmt = PIX_FMT_GBRP10;
2550                 } else
2551                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2552             } else if (CHROMA422)
2553                 s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2554             else
2555                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2556             break;
2557         case 12:
2558             if (CHROMA444) {
2559                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2560                     s->avctx->pix_fmt = PIX_FMT_GBRP12;
2561                 } else
2562                     s->avctx->pix_fmt = PIX_FMT_YUV444P12;
2563             } else if (CHROMA422)
2564                 s->avctx->pix_fmt = PIX_FMT_YUV422P12;
2565             else
2566                 s->avctx->pix_fmt = PIX_FMT_YUV420P12;
2567             break;
2568         case 14:
2569             if (CHROMA444) {
2570                 if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2571                     s->avctx->pix_fmt = PIX_FMT_GBRP14;
2572                 } else
2573                     s->avctx->pix_fmt = PIX_FMT_YUV444P14;
2574             } else if (CHROMA422)
2575                 s->avctx->pix_fmt = PIX_FMT_YUV422P14;
2576             else
2577                 s->avctx->pix_fmt = PIX_FMT_YUV420P14;
2578             break;
2579         case 8:
2580             if (CHROMA444) {
2581                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P
2582                                                                                   : PIX_FMT_YUV444P;
2583                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2584                         s->avctx->pix_fmt = PIX_FMT_GBR24P;
2585                         av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
2586                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
2587                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
2588                     }
2589             } else if (CHROMA422) {
2590                 s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P
2591                                                                               : PIX_FMT_YUV422P;
2592             } else {
2593                 s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2594                                                          s->avctx->codec->pix_fmts ?
2595                                                          s->avctx->codec->pix_fmts :
2596                                                          s->avctx->color_range == AVCOL_RANGE_JPEG ?
2597                                                          hwaccel_pixfmt_list_h264_jpeg_420 :
2598                                                          ff_hwaccel_pixfmt_list_420);
2599             }
2600             break;
2601         default:
2602             av_log(s->avctx, AV_LOG_ERROR,
2603                    "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
2604             return AVERROR_INVALIDDATA;
2605         }
2606
2607         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id,
2608                                             s->avctx->pix_fmt);
2609
2610         if (ff_MPV_common_init(s) < 0) {
2611             av_log(h->s.avctx, AV_LOG_ERROR, "ff_MPV_common_init() failed.\n");
2612             return -1;
2613         }
2614         s->first_field = 0;
2615         h->prev_interlaced_frame = 1;
2616
2617         init_scan_tables(h);
2618         if (ff_h264_alloc_tables(h) < 0) {
2619             av_log(h->s.avctx, AV_LOG_ERROR,
2620                    "Could not allocate memory for h264\n");
2621             return AVERROR(ENOMEM);
2622         }
2623
2624         if (!HAVE_THREADS || !(s->avctx->active_thread_type & FF_THREAD_SLICE)) {
2625             if (context_init(h) < 0) {
2626                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2627                 return -1;
2628             }
2629         } else {
2630             for (i = 1; i < s->slice_context_count; i++) {
2631                 H264Context *c;
2632                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2633                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2634                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2635                 c->h264dsp     = h->h264dsp;
2636                 c->sps         = h->sps;
2637                 c->pps         = h->pps;
2638                 c->pixel_shift = h->pixel_shift;
2639                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
2640                 init_scan_tables(c);
2641                 clone_tables(c, h, i);
2642             }
2643
2644             for (i = 0; i < s->slice_context_count; i++)
2645                 if (context_init(h->thread_context[i]) < 0) {
2646                     av_log(h->s.avctx, AV_LOG_ERROR,
2647                            "context_init() failed.\n");
2648                     return -1;
2649                 }
2650         }
2651     }
2652
2653     if (h == h0 && h->dequant_coeff_pps != pps_id) {
2654         h->dequant_coeff_pps = pps_id;
2655         init_dequant_tables(h);
2656     }
2657
2658     h->frame_num = get_bits(&s->gb, h->sps.log2_max_frame_num);
2659
2660     h->mb_mbaff        = 0;
2661     h->mb_aff_frame    = 0;
2662     last_pic_structure = s0->picture_structure;
2663     last_pic_dropable  = s->dropable;
2664     s->dropable        = h->nal_ref_idc == 0;
2665     if (h->sps.frame_mbs_only_flag) {
2666         s->picture_structure = PICT_FRAME;
2667     } else {
2668         if (!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B) {
2669             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
2670             return -1;
2671         }
2672         if (get_bits1(&s->gb)) { // field_pic_flag
2673             s->picture_structure = PICT_TOP_FIELD + get_bits1(&s->gb); // bottom_field_flag
2674         } else {
2675             s->picture_structure = PICT_FRAME;
2676             h->mb_aff_frame      = h->sps.mb_aff;
2677         }
2678     }
2679     h->mb_field_decoding_flag = s->picture_structure != PICT_FRAME;
2680
2681     if (h0->current_slice != 0) {
2682         if (last_pic_structure != s->picture_structure ||
2683             last_pic_dropable  != s->dropable) {
2684             av_log(h->s.avctx, AV_LOG_ERROR,
2685                    "Changing field mode (%d -> %d) between slices is not allowed\n",
2686                    last_pic_structure, s->picture_structure);
2687             s->picture_structure = last_pic_structure;
2688             s->dropable          = last_pic_dropable;
2689             return AVERROR_INVALIDDATA;
2690         }
2691     } else {
2692         /* Shorten frame num gaps so we don't have to allocate reference
2693          * frames just to throw them away */
2694         if (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
2695             int unwrap_prev_frame_num = h->prev_frame_num;
2696             int max_frame_num         = 1 << h->sps.log2_max_frame_num;
2697
2698             if (unwrap_prev_frame_num > h->frame_num)
2699                 unwrap_prev_frame_num -= max_frame_num;
2700
2701             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2702                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2703                 if (unwrap_prev_frame_num < 0)
2704                     unwrap_prev_frame_num += max_frame_num;
2705
2706                 h->prev_frame_num = unwrap_prev_frame_num;
2707             }
2708         }
2709
2710         /* See if we have a decoded first field looking for a pair...
2711          * Here, we're using that to see if we should mark previously
2712          * decode frames as "finished".
2713          * We have to do that before the "dummy" in-between frame allocation,
2714          * since that can modify s->current_picture_ptr. */
2715         if (s0->first_field) {
2716             assert(s0->current_picture_ptr);
2717             assert(s0->current_picture_ptr->f.data[0]);
2718             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2719
2720             /* Mark old field/frame as completed */
2721             if (!last_pic_dropable && s0->current_picture_ptr->owner2 == s0) {
2722                 ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2723                                           last_pic_structure == PICT_BOTTOM_FIELD);
2724             }
2725
2726             /* figure out if we have a complementary field pair */
2727             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2728                 /* Previous field is unmatched. Don't display it, but let it
2729                  * remain for reference if marked as such. */
2730                 if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
2731                     ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2732                                               last_pic_structure == PICT_TOP_FIELD);
2733                 }
2734             } else {
2735                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2736                     /* This and previous field were reference, but had
2737                      * different frame_nums. Consider this field first in
2738                      * pair. Throw away previous field except for reference
2739                      * purposes. */
2740                     if (!last_pic_dropable && last_pic_structure != PICT_FRAME) {
2741                         ff_thread_report_progress(&s0->current_picture_ptr->f, INT_MAX,
2742                                                   last_pic_structure == PICT_TOP_FIELD);
2743                     }
2744                 } else {
2745                     /* Second field in complementary pair */
2746                     if (!((last_pic_structure   == PICT_TOP_FIELD &&
2747                            s->picture_structure == PICT_BOTTOM_FIELD) ||
2748                           (last_pic_structure   == PICT_BOTTOM_FIELD &&
2749                            s->picture_structure == PICT_TOP_FIELD))) {
2750                         av_log(s->avctx, AV_LOG_ERROR,
2751                                "Invalid field mode combination %d/%d\n",
2752                                last_pic_structure, s->picture_structure);
2753                         s->picture_structure = last_pic_structure;
2754                         s->dropable          = last_pic_dropable;
2755                         return AVERROR_INVALIDDATA;
2756                     } else if (last_pic_dropable != s->dropable) {
2757                         av_log(s->avctx, AV_LOG_ERROR,
2758                                "Cannot combine reference and non-reference fields in the same frame\n");
2759                         av_log_ask_for_sample(s->avctx, NULL);
2760                         s->picture_structure = last_pic_structure;
2761                         s->dropable          = last_pic_dropable;
2762                         return AVERROR_INVALIDDATA;
2763                     }
2764
2765                     /* Take ownership of this buffer. Note that if another thread owned
2766                      * the first field of this buffer, we're not operating on that pointer,
2767                      * so the original thread is still responsible for reporting progress
2768                      * on that first field (or if that was us, we just did that above).
2769                      * By taking ownership, we assign responsibility to ourselves to
2770                      * report progress on the second field. */
2771                     s0->current_picture_ptr->owner2 = s0;
2772                 }
2773             }
2774         }
2775
2776         while (h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0 &&
2777                h->frame_num != (h->prev_frame_num + 1) % (1 << h->sps.log2_max_frame_num)) {
2778             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2779             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n",
2780                    h->frame_num, h->prev_frame_num);
2781             if (ff_h264_frame_start(h) < 0)
2782                 return -1;
2783             h->prev_frame_num++;
2784             h->prev_frame_num %= 1 << h->sps.log2_max_frame_num;
2785             s->current_picture_ptr->frame_num = h->prev_frame_num;
2786             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 0);
2787             ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX, 1);
2788             ff_generate_sliding_window_mmcos(h);
2789             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2790                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2791                 return AVERROR_INVALIDDATA;
2792             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2793              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2794              * about there being no actual duplicates.
2795              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2796              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2797              * be fixed. */
2798             if (h->short_ref_count) {
2799                 if (prev) {
2800                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2801                                   (const uint8_t **)prev->f.data, prev->f.linesize,
2802                                   s->avctx->pix_fmt, s->mb_width * 16, s->mb_height * 16);
2803                     h->short_ref[0]->poc = prev->poc + 2;
2804                 }
2805                 h->short_ref[0]->frame_num = h->prev_frame_num;
2806             }
2807         }
2808
2809         /* See if we have a decoded first field looking for a pair...
2810          * We're using that to see whether to continue decoding in that
2811          * frame, or to allocate a new one. */
2812         if (s0->first_field) {
2813             assert(s0->current_picture_ptr);
2814             assert(s0->current_picture_ptr->f.data[0]);
2815             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2816
2817             /* figure out if we have a complementary field pair */
2818             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2819                 /* Previous field is unmatched. Don't display it, but let it
2820                  * remain for reference if marked as such. */
2821                 s0->current_picture_ptr = NULL;
2822                 s0->first_field         = FIELD_PICTURE;
2823             } else {
2824                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2825                     ff_thread_report_progress((AVFrame*)s0->current_picture_ptr, INT_MAX,
2826                                               s0->picture_structure==PICT_BOTTOM_FIELD);
2827                     /* This and the previous field had different frame_nums.
2828                      * Consider this field first in pair. Throw away previous
2829                      * one except for reference purposes. */
2830                     s0->first_field         = 1;
2831                     s0->current_picture_ptr = NULL;
2832                 } else {
2833                     /* Second field in complementary pair */
2834                     s0->first_field = 0;
2835                 }
2836             }
2837         } else {
2838             /* Frame or first field in a potentially complementary pair */
2839             assert(!s0->current_picture_ptr);
2840             s0->first_field = FIELD_PICTURE;
2841         }
2842
2843         if (!FIELD_PICTURE || s0->first_field) {
2844             if (ff_h264_frame_start(h) < 0) {
2845                 s0->first_field = 0;
2846                 return -1;
2847             }
2848         } else {
2849             ff_release_unused_pictures(s, 0);
2850         }
2851     }
2852     if (h != h0)
2853         clone_slice(h, h0);
2854
2855     s->current_picture_ptr->frame_num = h->frame_num; // FIXME frame_num cleanup
2856
2857     assert(s->mb_num == s->mb_width * s->mb_height);
2858     if (first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2859         first_mb_in_slice >= s->mb_num) {
2860         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2861         return -1;
2862     }
2863     s->resync_mb_x = s->mb_x =  first_mb_in_slice % s->mb_width;
2864     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2865     if (s->picture_structure == PICT_BOTTOM_FIELD)
2866         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2867     assert(s->mb_y < s->mb_height);
2868
2869     if (s->picture_structure == PICT_FRAME) {
2870         h->curr_pic_num = h->frame_num;
2871         h->max_pic_num  = 1 << h->sps.log2_max_frame_num;
2872     } else {
2873         h->curr_pic_num = 2 * h->frame_num + 1;
2874         h->max_pic_num  = 1 << (h->sps.log2_max_frame_num + 1);
2875     }
2876
2877     if (h->nal_unit_type == NAL_IDR_SLICE)
2878         get_ue_golomb(&s->gb); /* idr_pic_id */
2879
2880     if (h->sps.poc_type == 0) {
2881         h->poc_lsb = get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2882
2883         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
2884             h->delta_poc_bottom = get_se_golomb(&s->gb);
2885     }
2886
2887     if (h->sps.poc_type == 1 && !h->sps.delta_pic_order_always_zero_flag) {
2888         h->delta_poc[0] = get_se_golomb(&s->gb);
2889
2890         if (h->pps.pic_order_present == 1 && s->picture_structure == PICT_FRAME)
2891             h->delta_poc[1] = get_se_golomb(&s->gb);
2892     }
2893
2894     init_poc(h);
2895
2896     if (h->pps.redundant_pic_cnt_present)
2897         h->redundant_pic_count = get_ue_golomb(&s->gb);
2898
2899     // set defaults, might be overridden a few lines later
2900     h->ref_count[0] = h->pps.ref_count[0];
2901     h->ref_count[1] = h->pps.ref_count[1];
2902
2903     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
2904         unsigned max[2];
2905         max[0] = max[1] = s->picture_structure == PICT_FRAME ? 15 : 31;
2906
2907         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
2908             h->direct_spatial_mv_pred = get_bits1(&s->gb);
2909         num_ref_idx_active_override_flag = get_bits1(&s->gb);
2910
2911         if (num_ref_idx_active_override_flag) {
2912             h->ref_count[0] = get_ue_golomb(&s->gb) + 1;
2913             if (h->slice_type_nos == AV_PICTURE_TYPE_B)
2914                 h->ref_count[1] = get_ue_golomb(&s->gb) + 1;
2915             else
2916                 // full range is spec-ok in this case, even for frames
2917                 max[1] = 31;
2918         }
2919
2920         if (h->ref_count[0]-1 > max[0] || h->ref_count[1]-1 > max[1]){
2921             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow %u > %u or %u > %u\n", h->ref_count[0]-1, max[0], h->ref_count[1]-1, max[1]);
2922             h->ref_count[0] = h->ref_count[1] = 1;
2923             return AVERROR_INVALIDDATA;
2924         }
2925
2926         if (h->slice_type_nos == AV_PICTURE_TYPE_B)
2927             h->list_count = 2;
2928         else
2929             h->list_count = 1;
2930     } else
2931         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
2932
2933     if (!default_ref_list_done)
2934         ff_h264_fill_default_ref_list(h);
2935
2936     if (h->slice_type_nos != AV_PICTURE_TYPE_I &&
2937         ff_h264_decode_ref_pic_list_reordering(h) < 0) {
2938         h->ref_count[1] = h->ref_count[0] = 0;
2939         return -1;
2940     }
2941
2942     if (h->slice_type_nos != AV_PICTURE_TYPE_I) {
2943         s->last_picture_ptr = &h->ref_list[0][0];
2944         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2945     }
2946     if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
2947         s->next_picture_ptr = &h->ref_list[1][0];
2948         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2949     }
2950
2951     if ((h->pps.weighted_pred && h->slice_type_nos == AV_PICTURE_TYPE_P) ||
2952         (h->pps.weighted_bipred_idc == 1 &&
2953          h->slice_type_nos == AV_PICTURE_TYPE_B))
2954         pred_weight_table(h);
2955     else if (h->pps.weighted_bipred_idc == 2 &&
2956              h->slice_type_nos == AV_PICTURE_TYPE_B) {
2957         implicit_weight_table(h, -1);
2958     } else {
2959         h->use_weight = 0;
2960         for (i = 0; i < 2; i++) {
2961             h->luma_weight_flag[i]   = 0;
2962             h->chroma_weight_flag[i] = 0;
2963         }
2964     }
2965
2966     if (h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2967         (s->avctx->err_recognition & AV_EF_EXPLODE))
2968         return AVERROR_INVALIDDATA;
2969
2970     if (FRAME_MBAFF) {
2971         ff_h264_fill_mbaff_ref_list(h);
2972
2973         if (h->pps.weighted_bipred_idc == 2 && h->slice_type_nos == AV_PICTURE_TYPE_B) {
2974             implicit_weight_table(h, 0);
2975             implicit_weight_table(h, 1);
2976         }
2977     }
2978
2979     if (h->slice_type_nos == AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2980         ff_h264_direct_dist_scale_factor(h);
2981     ff_h264_direct_ref_list_init(h);
2982
2983     if (h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac) {
2984         tmp = get_ue_golomb_31(&s->gb);
2985         if (tmp > 2) {
2986             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2987             return -1;
2988         }
2989         h->cabac_init_idc = tmp;
2990     }
2991
2992     h->last_qscale_diff = 0;
2993     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2994     if (tmp > 51 + 6 * (h->sps.bit_depth_luma - 8)) {
2995         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2996         return -1;
2997     }
2998     s->qscale       = tmp;
2999     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3000     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3001     // FIXME qscale / qp ... stuff
3002     if (h->slice_type == AV_PICTURE_TYPE_SP)
3003         get_bits1(&s->gb); /* sp_for_switch_flag */
3004     if (h->slice_type == AV_PICTURE_TYPE_SP ||
3005         h->slice_type == AV_PICTURE_TYPE_SI)
3006         get_se_golomb(&s->gb); /* slice_qs_delta */
3007
3008     h->deblocking_filter     = 1;
3009     h->slice_alpha_c0_offset = 52;
3010     h->slice_beta_offset     = 52;
3011     if (h->pps.deblocking_filter_parameters_present) {
3012         tmp = get_ue_golomb_31(&s->gb);
3013         if (tmp > 2) {
3014             av_log(s->avctx, AV_LOG_ERROR,
3015                    "deblocking_filter_idc %u out of range\n", tmp);
3016             return -1;
3017         }
3018         h->deblocking_filter = tmp;
3019         if (h->deblocking_filter < 2)
3020             h->deblocking_filter ^= 1;  // 1<->0
3021
3022         if (h->deblocking_filter) {
3023             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3024             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3025             if (h->slice_alpha_c0_offset > 104U ||
3026                 h->slice_beta_offset     > 104U) {
3027                 av_log(s->avctx, AV_LOG_ERROR,
3028                        "deblocking filter parameters %d %d out of range\n",
3029                        h->slice_alpha_c0_offset, h->slice_beta_offset);
3030                 return -1;
3031             }
3032         }
3033     }
3034
3035     if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
3036         (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY &&
3037          h->slice_type_nos != AV_PICTURE_TYPE_I) ||
3038         (s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  &&
3039          h->slice_type_nos == AV_PICTURE_TYPE_B) ||
3040         (s->avctx->skip_loop_filter >= AVDISCARD_NONREF &&
3041          h->nal_ref_idc == 0))
3042         h->deblocking_filter = 0;
3043
3044     if (h->deblocking_filter == 1 && h0->max_contexts > 1) {
3045         if (s->avctx->flags2 & CODEC_FLAG2_FAST) {
3046             /* Cheat slightly for speed:
3047              * Do not bother to deblock across slices. */
3048             h->deblocking_filter = 2;
3049         } else {
3050             h0->max_contexts = 1;
3051             if (!h0->single_decode_warning) {
3052                 av_log(s->avctx, AV_LOG_INFO,
3053                        "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3054                 h0->single_decode_warning = 1;
3055             }
3056             if (h != h0) {
3057                 av_log(h->s.avctx, AV_LOG_ERROR,
3058                        "Deblocking switched inside frame.\n");
3059                 return 1;
3060             }
3061         }
3062     }
3063     h->qp_thresh = 15 + 52 -
3064                    FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) -
3065                    FFMAX3(0,
3066                           h->pps.chroma_qp_index_offset[0],
3067                           h->pps.chroma_qp_index_offset[1]) +
3068                    6 * (h->sps.bit_depth_luma - 8);
3069
3070     h0->last_slice_type = slice_type;
3071     h->slice_num = ++h0->current_slice;
3072
3073     if (h->slice_num)
3074         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3075     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3076         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3077         && h->slice_num >= MAX_SLICES) {
3078         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3079         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3080     }
3081
3082     for (j = 0; j < 2; j++) {
3083         int id_list[16];
3084         int *ref2frm = h->ref2frm[h->slice_num & (MAX_SLICES - 1)][j];
3085         for (i = 0; i < 16; i++) {
3086             id_list[i] = 60;
3087             if (h->ref_list[j][i].f.data[0]) {
3088                 int k;
3089                 uint8_t *base = h->ref_list[j][i].f.base[0];
3090                 for (k = 0; k < h->short_ref_count; k++)
3091                     if (h->short_ref[k]->f.base[0] == base) {
3092                         id_list[i] = k;
3093                         break;
3094                     }
3095                 for (k = 0; k < h->long_ref_count; k++)
3096                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3097                         id_list[i] = h->short_ref_count + k;
3098                         break;
3099                     }
3100             }
3101         }
3102
3103         ref2frm[0]     =
3104             ref2frm[1] = -1;
3105         for (i = 0; i < 16; i++)
3106             ref2frm[i + 2] = 4 * id_list[i] +
3107                              (h->ref_list[j][i].f.reference & 3);
3108         ref2frm[18 + 0]     =
3109             ref2frm[18 + 1] = -1;
3110         for (i = 16; i < 48; i++)
3111             ref2frm[i + 4] = 4 * id_list[(i - 16) >> 1] +
3112                              (h->ref_list[j][i].f.reference & 3);
3113     }
3114
3115     // FIXME: fix draw_edges + PAFF + frame threads
3116     h->emu_edge_width  = (s->flags & CODEC_FLAG_EMU_EDGE ||
3117                           (!h->sps.frame_mbs_only_flag &&
3118                            s->avctx->active_thread_type))
3119                          ? 0 : 16;
3120     h->emu_edge_height = (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3121
3122     if (s->avctx->debug & FF_DEBUG_PICT_INFO) {
3123         av_log(h->s.avctx, AV_LOG_DEBUG,
3124                "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3125                h->slice_num,
3126                (s->picture_structure == PICT_FRAME ? "F" : s->picture_structure == PICT_TOP_FIELD ? "T" : "B"),
3127                first_mb_in_slice,
3128                av_get_picture_type_char(h->slice_type),
3129                h->slice_type_fixed ? " fix" : "",
3130                h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3131                pps_id, h->frame_num,
3132                s->current_picture_ptr->field_poc[0],
3133                s->current_picture_ptr->field_poc[1],
3134                h->ref_count[0], h->ref_count[1],
3135                s->qscale,
3136                h->deblocking_filter,
3137                h->slice_alpha_c0_offset / 2 - 26, h->slice_beta_offset / 2 - 26,
3138                h->use_weight,
3139                h->use_weight == 1 && h->use_weight_chroma ? "c" : "",
3140                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : "");
3141     }
3142
3143     return 0;
3144 }
3145
3146 int ff_h264_get_slice_type(const H264Context *h)
3147 {
3148     switch (h->slice_type) {
3149     case AV_PICTURE_TYPE_P:
3150         return 0;
3151     case AV_PICTURE_TYPE_B:
3152         return 1;
3153     case AV_PICTURE_TYPE_I:
3154         return 2;
3155     case AV_PICTURE_TYPE_SP:
3156         return 3;
3157     case AV_PICTURE_TYPE_SI:
3158         return 4;
3159     default:
3160         return -1;
3161     }
3162 }
3163
3164 static av_always_inline void fill_filter_caches_inter(H264Context *h,
3165                                                       MpegEncContext *const s,
3166                                                       int mb_type, int top_xy,
3167                                                       int left_xy[LEFT_MBS],
3168                                                       int top_type,
3169                                                       int left_type[LEFT_MBS],
3170                                                       int mb_xy, int list)
3171 {
3172     int b_stride = h->b_stride;
3173     int16_t(*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3174     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3175     if (IS_INTER(mb_type) || IS_DIRECT(mb_type)) {
3176         if (USES_LIST(top_type, list)) {
3177             const int b_xy  = h->mb2b_xy[top_xy] + 3 * b_stride;
3178             const int b8_xy = 4 * top_xy + 2;
3179             int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_table[top_xy] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3180             AV_COPY128(mv_dst - 1 * 8, s->current_picture.f.motion_val[list][b_xy + 0]);
3181             ref_cache[0 - 1 * 8] =
3182             ref_cache[1 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3183             ref_cache[2 - 1 * 8] =
3184             ref_cache[3 - 1 * 8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3185         } else {
3186             AV_ZERO128(mv_dst - 1 * 8);
3187             AV_WN32A(&ref_cache[0 - 1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3188         }
3189
3190         if (!IS_INTERLACED(mb_type ^ left_type[LTOP])) {
3191             if (USES_LIST(left_type[LTOP], list)) {
3192                 const int b_xy  = h->mb2b_xy[left_xy[LTOP]] + 3;
3193                 const int b8_xy = 4 * left_xy[LTOP] + 1;
3194                 int (*ref2frm)[64] =(void*)( h->ref2frm[h->slice_table[left_xy[LTOP]] & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3195                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride * 0]);
3196                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride * 1]);
3197                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride * 2]);
3198                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride * 3]);
3199                 ref_cache[-1 +  0] =
3200                 ref_cache[-1 +  8] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 0]];
3201                 ref_cache[-1 + 16] =
3202                 ref_cache[-1 + 24] = ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2 * 1]];
3203             } else {
3204                 AV_ZERO32(mv_dst - 1 +  0);
3205                 AV_ZERO32(mv_dst - 1 +  8);
3206                 AV_ZERO32(mv_dst - 1 + 16);
3207                 AV_ZERO32(mv_dst - 1 + 24);
3208                 ref_cache[-1 +  0] =
3209                 ref_cache[-1 +  8] =
3210                 ref_cache[-1 + 16] =
3211                 ref_cache[-1 + 24] = LIST_NOT_USED;
3212             }
3213         }
3214     }
3215
3216     if (!USES_LIST(mb_type, list)) {
3217         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0, 0), 4);
3218         AV_WN32A(&ref_cache[0 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3219         AV_WN32A(&ref_cache[1 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3220         AV_WN32A(&ref_cache[2 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3221         AV_WN32A(&ref_cache[3 * 8], ((LIST_NOT_USED) & 0xFF) * 0x01010101u);
3222         return;
3223     }
3224
3225     {
3226         int8_t *ref = &s->current_picture.f.ref_index[list][4 * mb_xy];
3227         int (*ref2frm)[64] = (void*)(h->ref2frm[h->slice_num & (MAX_SLICES - 1)][0] + (MB_MBAFF ? 20 : 2));
3228         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]], ref2frm[list][ref[1]]) & 0x00FF00FF) * 0x0101;
3229         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]], ref2frm[list][ref[3]]) & 0x00FF00FF) * 0x0101;
3230         AV_WN32A(&ref_cache[0 * 8], ref01);
3231         AV_WN32A(&ref_cache[1 * 8], ref01);
3232         AV_WN32A(&ref_cache[2 * 8], ref23);
3233         AV_WN32A(&ref_cache[3 * 8], ref23);
3234     }
3235
3236     {
3237         int16_t(*mv_src)[2] = &s->current_picture.f.motion_val[list][4 * s->mb_x + 4 * s->mb_y * b_stride];
3238         AV_COPY128(mv_dst + 8 * 0, mv_src + 0 * b_stride);
3239         AV_COPY128(mv_dst + 8 * 1, mv_src + 1 * b_stride);
3240         AV_COPY128(mv_dst + 8 * 2, mv_src + 2 * b_stride);
3241         AV_COPY128(mv_dst + 8 * 3, mv_src + 3 * b_stride);
3242     }
3243 }
3244
3245 /**
3246  *
3247  * @return non zero if the loop filter can be skipped
3248  */
3249 static int fill_filter_caches(H264Context *h, int mb_type)
3250 {
3251     MpegEncContext *const s = &h->s;
3252     const int mb_xy = h->mb_xy;
3253     int top_xy, left_xy[LEFT_MBS];
3254     int top_type, left_type[LEFT_MBS];
3255     uint8_t *nnz;
3256     uint8_t *nnz_cache;
3257
3258     top_xy = mb_xy - (s->mb_stride << MB_FIELD);
3259
3260     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3261      * stuff, I can't imagine that these complex rules are worth it. */
3262
3263     left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
3264     if (FRAME_MBAFF) {
3265         const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3266         const int curr_mb_field_flag = IS_INTERLACED(mb_type);
3267         if (s->mb_y & 1) {
3268             if (left_mb_field_flag != curr_mb_field_flag)
3269                 left_xy[LTOP] -= s->mb_stride;
3270         } else {
3271             if (curr_mb_field_flag)
3272                 top_xy += s->mb_stride &
3273                     (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3274             if (left_mb_field_flag != curr_mb_field_flag)
3275                 left_xy[LBOT] += s->mb_stride;
3276         }
3277     }
3278
3279     h->top_mb_xy        = top_xy;
3280     h->left_mb_xy[LTOP] = left_xy[LTOP];
3281     h->left_mb_xy[LBOT] = left_xy[LBOT];
3282     {
3283         /* For sufficiently low qp, filtering wouldn't do anything.
3284          * This is a conservative estimate: could also check beta_offset
3285          * and more accurate chroma_qp. */
3286         int qp_thresh = h->qp_thresh; // FIXME strictly we should store qp_thresh for each mb of a slice
3287         int qp        = s->current_picture.f.qscale_table[mb_xy];
3288         if (qp <= qp_thresh &&
3289             (left_xy[LTOP] < 0 ||
3290              ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh) &&
3291             (top_xy < 0 ||
3292              ((qp + s->current_picture.f.qscale_table[top_xy] + 1) >> 1) <= qp_thresh)) {
3293             if (!FRAME_MBAFF)
3294                 return 1;
3295             if ((left_xy[LTOP] < 0 ||
3296                  ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]] + 1) >> 1) <= qp_thresh) &&
3297                 (top_xy < s->mb_stride ||
3298                  ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3299                 return 1;
3300         }
3301     }
3302
3303     top_type        = s->current_picture.f.mb_type[top_xy];
3304     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3305     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3306     if (h->deblocking_filter == 2) {
3307         if (h->slice_table[top_xy] != h->slice_num)
3308             top_type = 0;
3309         if (h->slice_table[left_xy[LBOT]] != h->slice_num)
3310             left_type[LTOP] = left_type[LBOT] = 0;
3311     } else {
3312         if (h->slice_table[top_xy] == 0xFFFF)
3313             top_type = 0;
3314         if (h->slice_table[left_xy[LBOT]] == 0xFFFF)
3315             left_type[LTOP] = left_type[LBOT] = 0;
3316     }
3317     h->top_type        = top_type;
3318     h->left_type[LTOP] = left_type[LTOP];
3319     h->left_type[LBOT] = left_type[LBOT];
3320
3321     if (IS_INTRA(mb_type))
3322         return 0;
3323
3324     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3325                              top_type, left_type, mb_xy, 0);
3326     if (h->list_count == 2)
3327         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy,
3328                                  top_type, left_type, mb_xy, 1);
3329
3330     nnz       = h->non_zero_count[mb_xy];
3331     nnz_cache = h->non_zero_count_cache;
3332     AV_COPY32(&nnz_cache[4 + 8 * 1], &nnz[0]);
3333     AV_COPY32(&nnz_cache[4 + 8 * 2], &nnz[4]);
3334     AV_COPY32(&nnz_cache[4 + 8 * 3], &nnz[8]);
3335     AV_COPY32(&nnz_cache[4 + 8 * 4], &nnz[12]);
3336     h->cbp = h->cbp_table[mb_xy];
3337
3338     if (top_type) {
3339         nnz = h->non_zero_count[top_xy];
3340         AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[3 * 4]);
3341     }
3342
3343     if (left_type[LTOP]) {
3344         nnz = h->non_zero_count[left_xy[LTOP]];
3345         nnz_cache[3 + 8 * 1] = nnz[3 + 0 * 4];
3346         nnz_cache[3 + 8 * 2] = nnz[3 + 1 * 4];
3347         nnz_cache[3 + 8 * 3] = nnz[3 + 2 * 4];
3348         nnz_cache[3 + 8 * 4] = nnz[3 + 3 * 4];
3349     }
3350
3351     /* CAVLC 8x8dct requires NNZ values for residual decoding that differ
3352      * from what the loop filter needs */
3353     if (!CABAC && h->pps.transform_8x8_mode) {
3354         if (IS_8x8DCT(top_type)) {
3355             nnz_cache[4 + 8 * 0]     =
3356                 nnz_cache[5 + 8 * 0] = (h->cbp_table[top_xy] & 0x4000) >> 12;
3357             nnz_cache[6 + 8 * 0]     =
3358                 nnz_cache[7 + 8 * 0] = (h->cbp_table[top_xy] & 0x8000) >> 12;
3359         }
3360         if (IS_8x8DCT(left_type[LTOP])) {
3361             nnz_cache[3 + 8 * 1]     =
3362                 nnz_cache[3 + 8 * 2] = (h->cbp_table[left_xy[LTOP]] & 0x2000) >> 12; // FIXME check MBAFF
3363         }
3364         if (IS_8x8DCT(left_type[LBOT])) {
3365             nnz_cache[3 + 8 * 3]     =
3366                 nnz_cache[3 + 8 * 4] = (h->cbp_table[left_xy[LBOT]] & 0x8000) >> 12; // FIXME check MBAFF
3367         }
3368
3369         if (IS_8x8DCT(mb_type)) {
3370             nnz_cache[scan8[0]] =
3371             nnz_cache[scan8[1]] =
3372             nnz_cache[scan8[2]] =
3373             nnz_cache[scan8[3]] = (h->cbp & 0x1000) >> 12;
3374
3375             nnz_cache[scan8[0 + 4]] =
3376             nnz_cache[scan8[1 + 4]] =
3377             nnz_cache[scan8[2 + 4]] =
3378             nnz_cache[scan8[3 + 4]] = (h->cbp & 0x2000) >> 12;
3379
3380             nnz_cache[scan8[0 + 8]] =
3381             nnz_cache[scan8[1 + 8]] =
3382             nnz_cache[scan8[2 + 8]] =
3383             nnz_cache[scan8[3 + 8]] = (h->cbp & 0x4000) >> 12;
3384
3385             nnz_cache[scan8[0 + 12]] =
3386             nnz_cache[scan8[1 + 12]] =
3387             nnz_cache[scan8[2 + 12]] =
3388             nnz_cache[scan8[3 + 12]] = (h->cbp & 0x8000) >> 12;
3389         }
3390     }
3391
3392     return 0;
3393 }
3394
3395 static void loop_filter(H264Context *h, int start_x, int end_x)
3396 {
3397     MpegEncContext *const s = &h->s;
3398     uint8_t *dest_y, *dest_cb, *dest_cr;
3399     int linesize, uvlinesize, mb_x, mb_y;
3400     const int end_mb_y       = s->mb_y + FRAME_MBAFF;
3401     const int old_slice_type = h->slice_type;
3402     const int pixel_shift    = h->pixel_shift;
3403     const int block_h        = 16 >> s->chroma_y_shift;
3404
3405     if (h->deblocking_filter) {
3406         for (mb_x = start_x; mb_x < end_x; mb_x++)
3407             for (mb_y = end_mb_y - FRAME_MBAFF; mb_y <= end_mb_y; mb_y++) {
3408                 int mb_xy, mb_type;
3409                 mb_xy         = h->mb_xy = mb_x + mb_y * s->mb_stride;
3410                 h->slice_num  = h->slice_table[mb_xy];
3411                 mb_type       = s->current_picture.f.mb_type[mb_xy];
3412                 h->list_count = h->list_counts[mb_xy];
3413
3414                 if (FRAME_MBAFF)
3415                     h->mb_mbaff               =
3416                     h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3417
3418                 s->mb_x = mb_x;
3419                 s->mb_y = mb_y;
3420                 dest_y  = s->current_picture.f.data[0] +
3421                           ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
3422                 dest_cb = s->current_picture.f.data[1] +
3423                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3424                           mb_y * s->uvlinesize * block_h;
3425                 dest_cr = s->current_picture.f.data[2] +
3426                           (mb_x << pixel_shift) * (8 << CHROMA444) +
3427                           mb_y * s->uvlinesize * block_h;
3428                 // FIXME simplify above
3429
3430                 if (MB_FIELD) {
3431                     linesize   = h->mb_linesize   = s->linesize   * 2;
3432                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3433                     if (mb_y & 1) { // FIXME move out of this function?
3434                         dest_y  -= s->linesize   * 15;
3435                         dest_cb -= s->uvlinesize * (block_h - 1);
3436                         dest_cr -= s->uvlinesize * (block_h - 1);
3437                     }
3438                 } else {
3439                     linesize   = h->mb_linesize   = s->linesize;
3440                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3441                 }
3442                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize,
3443                                  uvlinesize, 0);
3444                 if (fill_filter_caches(h, mb_type))
3445                     continue;
3446                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3447                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3448
3449                 if (FRAME_MBAFF) {
3450                     ff_h264_filter_mb(h, mb_x, mb_y, dest_y, dest_cb, dest_cr,
3451                                       linesize, uvlinesize);
3452                 } else {
3453                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb,
3454                                            dest_cr, linesize, uvlinesize);
3455                 }
3456             }
3457     }
3458     h->slice_type   = old_slice_type;
3459     s->mb_x         = end_x;
3460     s->mb_y         = end_mb_y - FRAME_MBAFF;
3461     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3462     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3463 }
3464
3465 static void predict_field_decoding_flag(H264Context *h)
3466 {
3467     MpegEncContext *const s = &h->s;
3468     const int mb_xy = s->mb_x + s->mb_y * s->mb_stride;
3469     int mb_type     = (h->slice_table[mb_xy - 1] == h->slice_num) ?
3470                       s->current_picture.f.mb_type[mb_xy - 1] :
3471                       (h->slice_table[mb_xy - s->mb_stride] == h->slice_num) ?
3472                       s->current_picture.f.mb_type[mb_xy - s->mb_stride] : 0;
3473     h->mb_mbaff     = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3474 }
3475
3476 /**
3477  * Draw edges and report progress for the last MB row.
3478  */
3479 static void decode_finish_row(H264Context *h)
3480 {
3481     MpegEncContext *const s = &h->s;
3482     int top            = 16 * (s->mb_y      >> FIELD_PICTURE);
3483     int pic_height     = 16 *  s->mb_height >> FIELD_PICTURE;
3484     int height         =  16      << FRAME_MBAFF;
3485     int deblock_border = (16 + 4) << FRAME_MBAFF;
3486
3487     if (h->deblocking_filter) {
3488         if ((top + height) >= pic_height)
3489             height += deblock_border;
3490         top -= deblock_border;
3491     }
3492
3493     if (top >= pic_height || (top + height) < h->emu_edge_height)
3494         return;
3495
3496     height = FFMIN(height, pic_height - top);
3497     if (top < h->emu_edge_height) {
3498         height = top + height;
3499         top    = 0;
3500     }
3501
3502     ff_draw_horiz_band(s, top, height);
3503
3504     if (s->dropable)
3505         return;
3506
3507     ff_thread_report_progress(&s->current_picture_ptr->f, top + height - 1,
3508                               s->picture_structure == PICT_BOTTOM_FIELD);
3509 }
3510
3511 static int decode_slice(struct AVCodecContext *avctx, void *arg)
3512 {
3513     H264Context *h = *(void **)arg;
3514     MpegEncContext *const s = &h->s;
3515     const int part_mask     = s->partitioned_frame ? (ER_AC_END | ER_AC_ERROR)
3516                                                    : 0x7F;
3517     int lf_x_start = s->mb_x;
3518
3519     s->mb_skip_run = -1;
3520
3521     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME ||
3522                     s->codec_id != CODEC_ID_H264 ||
3523                     (CONFIG_GRAY && (s->flags & CODEC_FLAG_GRAY));
3524
3525     if (h->pps.cabac) {
3526         /* realign */
3527         align_get_bits(&s->gb);
3528
3529         /* init cabac */
3530         ff_init_cabac_states(&h->cabac);
3531         ff_init_cabac_decoder(&h->cabac,
3532                               s->gb.buffer + get_bits_count(&s->gb) / 8,
3533                               (get_bits_left(&s->gb) + 7) / 8);
3534
3535         ff_h264_init_cabac_states(h);
3536
3537         for (;;) {
3538             // START_TIMER
3539             int ret = ff_h264_decode_mb_cabac(h);
3540             int eos;
3541             // STOP_TIMER("decode_mb_cabac")
3542
3543             if (ret >= 0)
3544                 ff_h264_hl_decode_mb(h);
3545
3546             // FIXME optimal? or let mb_decode decode 16x32 ?
3547             if (ret >= 0 && FRAME_MBAFF) {
3548                 s->mb_y++;
3549
3550                 ret = ff_h264_decode_mb_cabac(h);
3551
3552                 if (ret >= 0)
3553                     ff_h264_hl_decode_mb(h);
3554                 s->mb_y--;
3555             }
3556             eos = get_cabac_terminate(&h->cabac);
3557
3558             if ((s->workaround_bugs & FF_BUG_TRUNCATED) &&
3559                 h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3560                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3561                                 s->mb_y, ER_MB_END & part_mask);
3562                 if (s->mb_x >= lf_x_start)
3563                     loop_filter(h, lf_x_start, s->mb_x + 1);
3564                 return 0;
3565             }
3566             if (h->cabac.bytestream > h->cabac.bytestream_end + 2 )
3567                 av_log(h->s.avctx, AV_LOG_DEBUG, "bytestream overread %td\n", h->cabac.bytestream_end - h->cabac.bytestream);
3568             if (ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 4) {
3569                 av_log(h->s.avctx, AV_LOG_ERROR,
3570                        "error while decoding MB %d %d, bytestream (%td)\n",
3571                        s->mb_x, s->mb_y,
3572                        h->cabac.bytestream_end - h->cabac.bytestream);
3573                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3574                                 s->mb_y, ER_MB_ERROR & part_mask);
3575                 return -1;
3576             }
3577
3578             if (++s->mb_x >= s->mb_width) {
3579                 loop_filter(h, lf_x_start, s->mb_x);
3580                 s->mb_x = lf_x_start = 0;
3581                 decode_finish_row(h);
3582                 ++s->mb_y;
3583                 if (FIELD_OR_MBAFF_PICTURE) {
3584                     ++s->mb_y;
3585                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
3586                         predict_field_decoding_flag(h);
3587                 }
3588             }
3589
3590             if (eos || s->mb_y >= s->mb_height) {
3591                 tprintf(s->avctx, "slice end %d %d\n",
3592                         get_bits_count(&s->gb), s->gb.size_in_bits);
3593                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x - 1,
3594                                 s->mb_y, ER_MB_END & part_mask);
3595                 if (s->mb_x > lf_x_start)
3596                     loop_filter(h, lf_x_start, s->mb_x);
3597                 return 0;
3598             }
3599         }
3600     } else {
3601         for (;;) {
3602             int ret = ff_h264_decode_mb_cavlc(h);
3603
3604             if (ret >= 0)
3605                 ff_h264_hl_decode_mb(h);
3606
3607             // FIXME optimal? or let mb_decode decode 16x32 ?
3608             if (ret >= 0 && FRAME_MBAFF) {
3609                 s->mb_y++;
3610                 ret = ff_h264_decode_mb_cavlc(h);
3611
3612                 if (ret >= 0)
3613                     ff_h264_hl_decode_mb(h);
3614                 s->mb_y--;
3615             }
3616
3617             if (ret < 0) {
3618                 av_log(h->s.avctx, AV_LOG_ERROR,
3619                        "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3620                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3621                                 s->mb_y, ER_MB_ERROR & part_mask);
3622                 return -1;
3623             }
3624
3625             if (++s->mb_x >= s->mb_width) {
3626                 loop_filter(h, lf_x_start, s->mb_x);
3627                 s->mb_x = lf_x_start = 0;
3628                 decode_finish_row(h);
3629                 ++s->mb_y;
3630                 if (FIELD_OR_MBAFF_PICTURE) {
3631                     ++s->mb_y;
3632                     if (FRAME_MBAFF && s->mb_y < s->mb_height)
3633                         predict_field_decoding_flag(h);
3634                 }
3635                 if (s->mb_y >= s->mb_height) {
3636                     tprintf(s->avctx, "slice end %d %d\n",
3637                             get_bits_count(&s->gb), s->gb.size_in_bits);
3638
3639                     if (   get_bits_left(&s->gb) == 0
3640                         || get_bits_left(&s->gb) > 0 && !(s->avctx->err_recognition & AV_EF_AGGRESSIVE)) {
3641                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3642                                         s->mb_x - 1, s->mb_y,
3643                                         ER_MB_END & part_mask);
3644
3645                         return 0;
3646                     } else {
3647                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3648                                         s->mb_x, s->mb_y,
3649                                         ER_MB_END & part_mask);
3650
3651                         return -1;
3652                     }
3653                 }
3654             }
3655
3656             if (get_bits_left(&s->gb) <= 0 && s->mb_skip_run <= 0) {
3657                 tprintf(s->avctx, "slice end %d %d\n",
3658                         get_bits_count(&s->gb), s->gb.size_in_bits);
3659                 if (get_bits_left(&s->gb) == 0) {
3660                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y,
3661                                     s->mb_x - 1, s->mb_y,
3662                                     ER_MB_END & part_mask);
3663                     if (s->mb_x > lf_x_start)
3664                         loop_filter(h, lf_x_start, s->mb_x);
3665
3666                     return 0;
3667                 } else {
3668                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x,
3669                                     s->mb_y, ER_MB_ERROR & part_mask);
3670
3671                     return -1;
3672                 }
3673             }
3674         }
3675     }
3676 }
3677
3678 /**
3679  * Call decode_slice() for each context.
3680  *
3681  * @param h h264 master context
3682  * @param context_count number of contexts to execute
3683  */
3684 static int execute_decode_slices(H264Context *h, int context_count)
3685 {
3686     MpegEncContext *const s     = &h->s;
3687     AVCodecContext *const avctx = s->avctx;
3688     H264Context *hx;
3689     int i;
3690
3691     if (s->avctx->hwaccel ||
3692         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3693         return 0;
3694     if (context_count == 1) {
3695         return decode_slice(avctx, &h);
3696     } else {
3697         for (i = 1; i < context_count; i++) {
3698             hx                    = h->thread_context[i];
3699             hx->s.err_recognition = avctx->err_recognition;
3700             hx->s.error_count     = 0;
3701             hx->x264_build        = h->x264_build;
3702         }
3703
3704         avctx->execute(avctx, decode_slice, h->thread_context,
3705                        NULL, context_count, sizeof(void *));
3706
3707         /* pull back stuff from slices to master context */
3708         hx                   = h->thread_context[context_count - 1];
3709         s->mb_x              = hx->s.mb_x;
3710         s->mb_y              = hx->s.mb_y;
3711         s->dropable          = hx->s.dropable;
3712         s->picture_structure = hx->s.picture_structure;
3713         for (i = 1; i < context_count; i++)
3714             h->s.error_count += h->thread_context[i]->s.error_count;
3715     }
3716
3717     return 0;
3718 }
3719
3720 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
3721 {
3722     MpegEncContext *const s     = &h->s;
3723     AVCodecContext *const avctx = s->avctx;
3724     H264Context *hx; ///< thread context
3725     int buf_index;
3726     int context_count;
3727     int next_avc;
3728     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3729     int nals_needed = 0; ///< number of NALs that need decoding before the next frame thread starts
3730     int nal_index;
3731
3732     h->nal_unit_type= 0;
3733
3734     if(!s->slice_context_count)
3735          s->slice_context_count= 1;
3736     h->max_contexts = s->slice_context_count;
3737     if (!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3738         h->current_slice = 0;
3739         if (!s->first_field)
3740             s->current_picture_ptr = NULL;
3741         ff_h264_reset_sei(h);
3742     }
3743
3744     for (; pass <= 1; pass++) {
3745         buf_index     = 0;
3746         context_count = 0;
3747         next_avc      = h->is_avc ? 0 : buf_size;
3748         nal_index     = 0;
3749         for (;;) {
3750             int consumed;
3751             int dst_length;
3752             int bit_length;
3753             const uint8_t *ptr;
3754             int i, nalsize = 0;
3755             int err;
3756
3757             if (buf_index >= next_avc) {
3758                 if (buf_index >= buf_size - h->nal_length_size)
3759                     break;
3760                 nalsize = 0;
3761                 for (i = 0; i < h->nal_length_size; i++)
3762                     nalsize = (nalsize << 8) | buf[buf_index++];
3763                 if (nalsize <= 0 || nalsize > buf_size - buf_index) {
3764                     av_log(h->s.avctx, AV_LOG_ERROR,
3765                            "AVC: nal size %d\n", nalsize);
3766                     break;
3767                 }
3768                 next_avc = buf_index + nalsize;
3769             } else {
3770                 // start code prefix search
3771                 for (; buf_index + 3 < next_avc; buf_index++)
3772                     // This should always succeed in the first iteration.
3773                     if (buf[buf_index]     == 0 &&
3774                         buf[buf_index + 1] == 0 &&
3775                         buf[buf_index + 2] == 1)
3776                         break;
3777
3778                 if (buf_index + 3 >= buf_size)
3779                     break;
3780
3781                 buf_index += 3;
3782                 if (buf_index >= next_avc)
3783                     continue;
3784             }
3785
3786             hx = h->thread_context[context_count];
3787
3788             ptr = ff_h264_decode_nal(hx, buf + buf_index, &dst_length,
3789                                      &consumed, next_avc - buf_index);
3790             if (ptr == NULL || dst_length < 0) {
3791                 buf_index = -1;
3792                 goto end;
3793             }
3794             i = buf_index + consumed;
3795             if ((s->workaround_bugs & FF_BUG_AUTODETECT) && i + 3 < next_avc &&
3796                 buf[i]     == 0x00 && buf[i + 1] == 0x00 &&
3797                 buf[i + 2] == 0x01 && buf[i + 3] == 0xE0)
3798                 s->workaround_bugs |= FF_BUG_TRUNCATED;
3799
3800             if (!(s->workaround_bugs & FF_BUG_TRUNCATED))
3801                 while(dst_length > 0 && ptr[dst_length - 1] == 0)
3802                     dst_length--;
3803             bit_length = !dst_length ? 0
3804                                      : (8 * dst_length -
3805                                         decode_rbsp_trailing(h, ptr + dst_length - 1));
3806
3807             if (s->avctx->debug & FF_DEBUG_STARTCODE)
3808                 av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d pass %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length, pass);
3809
3810             if (h->is_avc && (nalsize != consumed) && nalsize)
3811                 av_log(h->s.avctx, AV_LOG_DEBUG,
3812                        "AVC: Consumed only %d bytes instead of %d\n",
3813                        consumed, nalsize);
3814
3815             buf_index += consumed;
3816             nal_index++;
3817
3818             if (pass == 0) {
3819                 /* packets can sometimes contain multiple PPS/SPS,
3820                  * e.g. two PAFF field pictures in one packet, or a demuxer
3821                  * which splits NALs strangely if so, when frame threading we
3822                  * can't start the next thread until we've read all of them */
3823                 switch (hx->nal_unit_type) {
3824                 case NAL_SPS:
3825                 case NAL_PPS:
3826                     nals_needed = nal_index;
3827                     break;
3828                 case NAL_IDR_SLICE:
3829                 case NAL_SLICE:
3830                     init_get_bits(&hx->s.gb, ptr, bit_length);
3831                     if (!get_ue_golomb(&hx->s.gb))
3832                         nals_needed = nal_index;
3833                 }
3834                 continue;
3835             }
3836
3837             // FIXME do not discard SEI id
3838             if (avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc == 0)
3839                 continue;
3840
3841 again:
3842             err = 0;
3843             switch (hx->nal_unit_type) {
3844             case NAL_IDR_SLICE:
3845                 if (h->nal_unit_type != NAL_IDR_SLICE) {
3846                     av_log(h->s.avctx, AV_LOG_ERROR,
3847                            "Invalid mix of idr and non-idr slices\n");
3848                     buf_index = -1;
3849                     goto end;
3850                 }
3851                 idr(h); // FIXME ensure we don't lose some frames if there is reordering
3852             case NAL_SLICE:
3853                 init_get_bits(&hx->s.gb, ptr, bit_length);
3854                 hx->intra_gb_ptr        =
3855                     hx->inter_gb_ptr    = &hx->s.gb;
3856                 hx->s.data_partitioning = 0;
3857
3858                 if ((err = decode_slice_header(hx, h)))
3859                     break;
3860
3861                 if (   h->sei_recovery_frame_cnt >= 0
3862                     && (   h->recovery_frame<0
3863                         || ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt)) {
3864                     h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
3865                                         (1 << h->sps.log2_max_frame_num);
3866                 }
3867
3868                 s->current_picture_ptr->f.key_frame |=
3869                         (hx->nal_unit_type == NAL_IDR_SLICE);
3870
3871                 if (h->recovery_frame == h->frame_num) {
3872                     s->current_picture_ptr->sync |= 1;
3873                     h->recovery_frame = -1;
3874                 }
3875
3876                 h->sync |= !!s->current_picture_ptr->f.key_frame;
3877                 h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
3878                 s->current_picture_ptr->sync |= h->sync;
3879
3880                 if (h->current_slice == 1) {
3881                     if (!(s->flags2 & CODEC_FLAG2_CHUNKS))
3882                         decode_postinit(h, nal_index >= nals_needed);
3883
3884                     if (s->avctx->hwaccel &&
3885                         s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3886                         return -1;
3887                     if (CONFIG_H264_VDPAU_DECODER &&
3888                         s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU)
3889                         ff_vdpau_h264_picture_start(s);
3890                 }
3891
3892                 if (hx->redundant_pic_count == 0 &&
3893                     (avctx->skip_frame < AVDISCARD_NONREF ||
3894                      hx->nal_ref_idc) &&
3895                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
3896                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
3897                     (avctx->skip_frame < AVDISCARD_NONKEY ||
3898                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
3899                     avctx->skip_frame < AVDISCARD_ALL) {
3900                     if (avctx->hwaccel) {
3901                         if (avctx->hwaccel->decode_slice(avctx,
3902                                                          &buf[buf_index - consumed],
3903                                                          consumed) < 0)
3904                             return -1;
3905                     } else if (CONFIG_H264_VDPAU_DECODER &&
3906                                s->avctx->codec->capabilities & CODEC_CAP_HWACCEL_VDPAU) {
3907                         static const uint8_t start_code[] = {
3908                             0x00, 0x00, 0x01 };
3909                         ff_vdpau_add_data_chunk(s, start_code,
3910                                                 sizeof(start_code));
3911                         ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed],
3912                                                 consumed);
3913                     } else
3914                         context_count++;
3915                 }
3916                 break;
3917             case NAL_DPA:
3918                 init_get_bits(&hx->s.gb, ptr, bit_length);
3919                 hx->intra_gb_ptr =
3920                 hx->inter_gb_ptr = NULL;
3921
3922                 if ((err = decode_slice_header(hx, h)) < 0)
3923                     break;
3924
3925                 hx->s.data_partitioning = 1;
3926                 break;
3927             case NAL_DPB:
3928                 init_get_bits(&hx->intra_gb, ptr, bit_length);
3929                 hx->intra_gb_ptr = &hx->intra_gb;
3930                 break;
3931             case NAL_DPC:
3932                 init_get_bits(&hx->inter_gb, ptr, bit_length);
3933                 hx->inter_gb_ptr = &hx->inter_gb;
3934
3935                 av_log(h->s.avctx, AV_LOG_ERROR, "Partitioned H.264 support is incomplete\n");
3936                 return AVERROR_PATCHWELCOME;
3937
3938                 if (hx->redundant_pic_count == 0 &&
3939                     hx->intra_gb_ptr &&
3940                     hx->s.data_partitioning &&
3941                     s->context_initialized &&
3942                     (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc) &&
3943                     (avctx->skip_frame < AVDISCARD_BIDIR  ||
3944                      hx->slice_type_nos != AV_PICTURE_TYPE_B) &&
3945                     (avctx->skip_frame < AVDISCARD_NONKEY ||
3946                      hx->slice_type_nos == AV_PICTURE_TYPE_I) &&
3947                     avctx->skip_frame < AVDISCARD_ALL)
3948                     context_count++;
3949                 break;
3950             case NAL_SEI:
3951                 init_get_bits(&s->gb, ptr, bit_length);
3952                 ff_h264_decode_sei(h);
3953                 break;
3954             case NAL_SPS:
3955                 init_get_bits(&s->gb, ptr, bit_length);
3956                 if (ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)) {
3957                     av_log(h->s.avctx, AV_LOG_DEBUG,
3958                            "SPS decoding failure, trying again with the complete NAL\n");
3959                     if (h->is_avc)
3960                         av_assert0(next_avc - buf_index + consumed == nalsize);
3961                     init_get_bits(&s->gb, &buf[buf_index + 1 - consumed],
3962                                   8*(next_avc - buf_index + consumed - 1));
3963                     ff_h264_decode_seq_parameter_set(h);
3964                 }
3965
3966                 if (s->flags & CODEC_FLAG_LOW_DELAY ||
3967                     (h->sps.bitstream_restriction_flag &&
3968                      !h->sps.num_reorder_frames))
3969                     s->low_delay = 1;
3970                 if (avctx->has_b_frames < 2)
3971                     avctx->has_b_frames = !s->low_delay;
3972                 break;
3973             case NAL_PPS:
3974                 init_get_bits(&s->gb, ptr, bit_length);
3975                 ff_h264_decode_picture_parameter_set(h, bit_length);
3976                 break;
3977             case NAL_AUD:
3978             case NAL_END_SEQUENCE:
3979             case NAL_END_STREAM:
3980             case NAL_FILLER_DATA:
3981             case NAL_SPS_EXT:
3982             case NAL_AUXILIARY_SLICE:
3983                 break;
3984             default:
3985                 av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n",
3986                        hx->nal_unit_type, bit_length);
3987             }
3988
3989             if (context_count == h->max_contexts) {
3990                 execute_decode_slices(h, context_count);
3991                 context_count = 0;
3992             }
3993
3994             if (err < 0)
3995                 av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3996             else if (err == 1) {
3997                 /* Slice could not be decoded in parallel mode, copy down
3998                  * NAL unit stuff to context 0 and restart. Note that
3999                  * rbsp_buffer is not transferred, but since we no longer
4000                  * run in parallel mode this should not be an issue. */
4001                 h->nal_unit_type = hx->nal_unit_type;
4002                 h->nal_ref_idc   = hx->nal_ref_idc;
4003                 hx               = h;
4004                 goto again;
4005             }
4006         }
4007     }
4008     if (context_count)
4009         execute_decode_slices(h, context_count);
4010
4011 end:
4012     /* clean up */
4013     if (s->current_picture_ptr && s->current_picture_ptr->owner2 == s &&
4014         !s->dropable) {
4015         ff_thread_report_progress(&s->current_picture_ptr->f, INT_MAX,
4016                                   s->picture_structure == PICT_BOTTOM_FIELD);
4017     }
4018
4019     return buf_index;
4020 }
4021
4022 /**
4023  * Return the number of bytes consumed for building the current frame.
4024  */
4025 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size)
4026 {
4027     if (pos == 0)
4028         pos = 1;          // avoid infinite loops (i doubt that is needed but ...)
4029     if (pos + 10 > buf_size)
4030         pos = buf_size;                   // oops ;)
4031
4032     return pos;
4033 }
4034
4035 static int decode_frame(AVCodecContext *avctx, void *data,
4036                         int *data_size, AVPacket *avpkt)
4037 {
4038     const uint8_t *buf = avpkt->data;
4039     int buf_size       = avpkt->size;
4040     H264Context *h     = avctx->priv_data;
4041     MpegEncContext *s  = &h->s;
4042     AVFrame *pict      = data;
4043     int buf_index      = 0;
4044     Picture *out;
4045     int i, out_idx;
4046
4047     s->flags  = avctx->flags;
4048     s->flags2 = avctx->flags2;
4049
4050     /* end of stream, output what is still in the buffers */
4051     if (buf_size == 0) {
4052  out:
4053
4054         s->current_picture_ptr = NULL;
4055
4056         // FIXME factorize this with the output code below
4057         out     = h->delayed_pic[0];
4058         out_idx = 0;
4059         for (i = 1;
4060              h->delayed_pic[i] &&
4061              !h->delayed_pic[i]->f.key_frame &&
4062              !h->delayed_pic[i]->mmco_reset;
4063              i++)
4064             if (h->delayed_pic[i]->poc < out->poc) {
4065                 out     = h->delayed_pic[i];
4066                 out_idx = i;
4067             }
4068
4069         for (i = out_idx; h->delayed_pic[i]; i++)
4070             h->delayed_pic[i] = h->delayed_pic[i + 1];
4071
4072         if (out) {
4073             *data_size = sizeof(AVFrame);
4074             *pict      = out->f;
4075         }
4076
4077         return buf_index;
4078     }
4079     if(h->is_avc && buf_size >= 9 && buf[0]==1 && buf[2]==0 && (buf[4]&0xFC)==0xFC && (buf[5]&0x1F) && buf[8]==0x67){
4080         int cnt= buf[5]&0x1f;
4081         const uint8_t *p= buf+6;
4082         while(cnt--){
4083             int nalsize= AV_RB16(p) + 2;
4084             if(nalsize > buf_size - (p-buf) || p[2]!=0x67)
4085                 goto not_extra;
4086             p += nalsize;
4087         }
4088         cnt = *(p++);
4089         if(!cnt)
4090             goto not_extra;
4091         while(cnt--){
4092             int nalsize= AV_RB16(p) + 2;
4093             if(nalsize > buf_size - (p-buf) || p[2]!=0x68)
4094                 goto not_extra;
4095             p += nalsize;
4096         }
4097
4098         return ff_h264_decode_extradata(h, buf, buf_size);
4099     }
4100 not_extra:
4101
4102     buf_index = decode_nal_units(h, buf, buf_size);
4103     if (buf_index < 0)
4104         return -1;
4105
4106     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4107         av_assert0(buf_index <= buf_size);
4108         goto out;
4109     }
4110
4111     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr) {
4112         if (avctx->skip_frame >= AVDISCARD_NONREF ||
4113             buf_size >= 4 && !memcmp("Q264", buf, 4))
4114             return buf_size;
4115         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4116         return -1;
4117     }
4118
4119     if (!(s->flags2 & CODEC_FLAG2_CHUNKS) ||
4120         (s->mb_y >= s->mb_height && s->mb_height)) {
4121         if (s->flags2 & CODEC_FLAG2_CHUNKS)
4122             decode_postinit(h, 1);
4123
4124         field_end(h, 0);
4125
4126         /* Wait for second field. */
4127         *data_size = 0;
4128         if (h->next_output_pic && (h->next_output_pic->sync || h->sync>1)) {
4129             *data_size = sizeof(AVFrame);
4130             *pict      = h->next_output_pic->f;
4131         }
4132     }
4133
4134     assert(pict->data[0] || !*data_size);
4135     ff_print_debug_info(s, pict);
4136     // printf("out %d\n", (int)pict->data[0]);
4137
4138     return get_consumed_bytes(s, buf_index, buf_size);
4139 }
4140
4141 av_cold void ff_h264_free_context(H264Context *h)
4142 {
4143     int i;
4144
4145     free_tables(h, 1); // FIXME cleanup init stuff perhaps
4146
4147     for (i = 0; i < MAX_SPS_COUNT; i++)
4148         av_freep(h->sps_buffers + i);
4149
4150     for (i = 0; i < MAX_PPS_COUNT; i++)
4151         av_freep(h->pps_buffers + i);
4152 }
4153
4154 static av_cold int h264_decode_end(AVCodecContext *avctx)
4155 {
4156     H264Context *h    = avctx->priv_data;
4157     MpegEncContext *s = &h->s;
4158
4159     ff_h264_remove_all_refs(h);
4160     ff_h264_free_context(h);
4161
4162     ff_MPV_common_end(s);
4163
4164     // memset(h, 0, sizeof(H264Context));
4165
4166     return 0;
4167 }
4168
4169 static const AVProfile profiles[] = {
4170     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4171     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4172     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4173     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4174     { FF_PROFILE_H264_HIGH,                 "High"                  },
4175     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4176     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4177     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4178     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4179     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4180     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4181     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4182     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4183     { FF_PROFILE_UNKNOWN },
4184 };
4185
4186 static const AVOption h264_options[] = {
4187     {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4188     {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4189     {NULL}
4190 };
4191
4192 static const AVClass h264_class = {
4193     "H264 Decoder",
4194     av_default_item_name,
4195     h264_options,
4196     LIBAVUTIL_VERSION_INT,
4197 };
4198
4199 static const AVClass h264_vdpau_class = {
4200     "H264 VDPAU Decoder",
4201     av_default_item_name,
4202     h264_options,
4203     LIBAVUTIL_VERSION_INT,
4204 };
4205
4206 AVCodec ff_h264_decoder = {
4207     .name                  = "h264",
4208     .type                  = AVMEDIA_TYPE_VIDEO,
4209     .id                    = CODEC_ID_H264,
4210     .priv_data_size        = sizeof(H264Context),
4211     .init                  = ff_h264_decode_init,
4212     .close                 = h264_decode_end,
4213     .decode                = decode_frame,
4214     .capabilities          = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 |
4215                              CODEC_CAP_DELAY | CODEC_CAP_SLICE_THREADS |
4216                              CODEC_CAP_FRAME_THREADS,
4217     .flush                 = flush_dpb,
4218     .long_name             = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4219     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4220     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4221     .profiles              = NULL_IF_CONFIG_SMALL(profiles),
4222     .priv_class            = &h264_class,
4223 };
4224
4225 #if CONFIG_H264_VDPAU_DECODER
4226 AVCodec ff_h264_vdpau_decoder = {
4227     .name           = "h264_vdpau",
4228     .type           = AVMEDIA_TYPE_VIDEO,
4229     .id             = CODEC_ID_H264,
4230     .priv_data_size = sizeof(H264Context),
4231     .init           = ff_h264_decode_init,
4232     .close          = h264_decode_end,
4233     .decode         = decode_frame,
4234     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4235     .flush          = flush_dpb,
4236     .long_name      = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4237     .pix_fmts       = (const enum PixelFormat[]) { PIX_FMT_VDPAU_H264,
4238                                                    PIX_FMT_NONE},
4239     .profiles       = NULL_IF_CONFIG_SMALL(profiles),
4240     .priv_class     = &h264_vdpau_class,
4241 };
4242 #endif